Challenges in medical additive manufacturing

Maureen Antonia Johanna Maria van Eijnatten

The studies presented in this thesis were conducted at the Department of Oral and Maxillofacial Surgery / Oral Pathology and the 3D Innovation Lab of the VU University Medical Center, Amsterdam, The Netherlands.

The printing of this thesis has been financially supported by the department of Oral and Maxillofacial Surgery / Oral Pathology of the VU University Medical Center and the VU University Amsterdam.

Cover, layout & printing: Off Page, Amsterdam

ISBN: 978-94-6182-846-0

© Maureen van Eijnatten, 2017

VRIJE UNIVERSITEIT

Challenges in medical additive manufacturing

ACADEMISCH PROEFSCHRIFT

ter verkrijging van de graad Doctor aan

de Vrije Universiteit Amsterdam,

op gezag van de rector magnificus

prof.dr. V. Subramaniam,

in het openbaar te verdedigen

ten overstaan van de promotiecommissie

van de Faculteit der Geneeskunde

op vrijdag 8 december 2017 om 9.45 uur

in de aula van de universiteit,

De Boelelaan 1105

door

Maureen Antonia Johanna Maria van Eijnatten

geboren te Zwolle

promotor: prof.dr. T. Forouzanfar

copromotor: dr. J.E.H. Wolff

TABLE OF CONTENTS

Chapter 1 General introduction 7

Chapter 2 A Novel Method of Orbital Floor Reconstruction Using 17 Virtual Planning, 3-Dimensional Printing, and Autologous Bone

Chapter 3 Influence of CT parameters on STL model accuracy 27

Chapter 4 Impact of prone, supine and oblique patient positioning on CBCT 43 image quality, contrast to noise ratio and figure of merit value in the maxillofacial region

Chapter 5 CT image segmentation methods for bone used in medical 61 additive manufacturing: A Literature Review

Chapter 6 The impact of manual threshold selection in 87 medical additive manufacturing

Chapter 7 Accuracy of MDCT and CBCT in three-dimensional 103 evaluation of the upper airway morphology

Chapter 8 The accuracy of ultrashort echo time MRI sequences for 117 medical additive manufacturing

Chapter 9 General discussion 133

Chapter 10 English summary & Dutch summary 141

Chapter 11 Acknowledgements 153

1 GENERAL INTRODUCTION

GE

NE

RA

L INTR

OD

UC

TION

9

1In the early 1980s, a revolutionary layer-by-layer manufacturing technique was developed

to create industrial prototype parts from three-dimensional (3D) computational data. This

novel approach was termed “rapid prototyping” (RP) and by the general public at large

as “3D printing”. The first commercial photopolymer 3D printer was patented by Charles

W. Hull in 1986 [1]. Since then, a plethora of different 3D printing technologies have

been developed that include selective laser sintering (SLS), fused deposition modelling

(FDM) and inkjet-based technologies [2]. Overtime, the term “rapid prototyping” has

been replaced by “additive manufacturing” (AM), a term that currently encompasses

a wide range of different technologies and applications [3]. AM technologies are currently

being used in many fields, including the automotive, aerospace and consumer goods

(customisation) industries and dental and medical science. According to the Wohlers

Report 2017, the AM industry grew by 17.4% in 2016 with a global market share of $6.1

billion. Furthermore, total sales volume is expected to reach $25.6 billion in 2021 [4].

The major advantages of AM technologies over traditional subtractive manufacturing

technologies are the possibilities to customise and create almost any complex geometry

in a completely automated manner [5]. Another advantage is the speed with which

a complex geometrical prototype or functional part can be manufactured when compared

to conventional subtractive manufacturing technologies.

Over the past decades, advances in image processing and graphical processing power

have extended the role of medical imaging far beyond traditional, two-dimensional

visualisation. Novel image processing techniques can convert a stack of two-dimensional

(2D) computed tomography (CT) or magnetic resonance (MR) images into a virtual 3D

model. Such 3D models can be used as a blueprint to manufacture medical constructs

using AM. Expiring patents and price cuts in medical 3D imaging and AM technologies

have sparked an increase in the implementation of medical AM constructs in clinical

settings [6].

Currently, there is a paradigm shift in medicine away from mass production towards

iterations and from one-size-fits-all to personalised treatments. In this context, AM constructs

have proven to be particularly valuable in the field of oral and maxillofacial surgery due

to the plethora of complex bony geometries found in the skull area. Moreover, only minor

dislocations in the facial area can have a profound impact on the function and aesthetic

appearance of patients. In such cases, patient-specific AM anatomical models have been

shown to improve communication between clinicians and patients [7]. In addition, AM

models are being used for the preoperative planning and simulation of complex surgery in

order to significantly improve the efficiency and outcome of the surgery [8]. Maxillofacial

surgeons are also routinely using patient-specific AM drill guides [9] and saw guides [10]

to translate a surgical plan into the operating room. To date, AM has been sporadically

used to create patient-specific implants of the mandible [11], orbita [12]–[14], zygoma [15],

cranium [16], tibia [17], and, although still in its infancy, scaffolds for tissue engineering

[18]. Lastly, AM models of (rare) anatomical abnormalities can, in certain circumstances,

GE

NE

RA

L INTR

OD

UC

TION

10

1replace cadaveric materials in medical education and training [19], [20], thereby avoiding

ethical controversies and reducing costs [21].

The current medical AM process comprises three different steps (Figure 1) [22]. The first

step is 3D image acquisition using either CT or MRI technologies. In clinical settings,

however, CT images are currently most often used due to their superior hard tissue

contrast, spatial resolution and lower cost. The resulting imaging data is subsequently

reconstructed in two-dimensional (2D) grey scale images that are saved in Digital Imaging

and Communications in Medicine (DICOM) file formats. The second step is image

processing, in which the tissue of interest (e.g., “bone”) is segmented and converted into

a virtual 3D surface model by means of triangulation [23]. The resulting 3D surface model

is saved as a Standard Tessellation Language (STL) model, which can then be used to

design patient-specific constructs using dedicated computer-aided design (CAD) software.

The third and final step in the medical AM process is the conversion of the designated STL

model into a G-code that subsequently controls the path of the printing nozzle during AM.

One major challenge faced in the current medical AM workflow (Figure 1) is the lack

of cohesive convergence between the medical imaging, computing and manufacturing

technologies. This lack of convergence is mainly due to the fact that no single company

provides all the necessary devices and services required for the medical AM process.

The large variety of different CT/MRI scanners, software packages, 3D printing techniques

and materials available on the market makes it difficult to select appropriate devices and

software for medical AM. Moreover, such devices and software packages are often not

certified for medical AM. As a result, concerns have subsequently been raised amongst

policymakers and regulators responsible for product safety. Manufacturers of imaging

devices, 3D printers and reverse engineering software companies should therefore

agree on technical norms for the medical AM process. This is particularly important in

the light of the new European Union medical device regulation (MDR) [24] that will require

standardisation and CE certification for the whole medical AM process from 2020 onwards.

With regard to the first step in the AM workflow (Figure 1, step 1), there is general

agreement amongst physicians and medical technicians that fan-beam CT technologies,

Figure 1. The three different steps required to create a patient-specific AM construct.

GE

NE

RA

L INTR

OD

UC

TION

11

1such as multi-detector row computed tomography (MDCT), offer the best images for

medical AM [25]. However, it still remains unclear which CT technology offers the most

accurate images for medical AM purposes. Moreover, to the best of my knowledge,

no standardised scan protocols have been specifically developed for medical AM. [26]

Currently, manufacturers are exploring new ways to improve CT image quality by using

multiple photon energies (dual-energy CT) to enhance tissue characterisation. Furthermore,

as opposed to conventional CT scanners, cone-beam computed tomography (CBCT)

scanners are currently gaining popularity, especially in small dental and maxillofacial

clinics. This surge in popularity is mainly due to the markedley lower puchsing costs,

acceptable image quality and low radiation dose when compared with conventional CT

scanners [27]. Recent studies have shown, however, that the partial CBCT gantry rotation

produces a non-uniform dose distribution [28] that can in certain cases affect CBCT image

quality. Therefore, CBCT scanning technology should not be routinely used to fabricate

patient-specific implants. Interestingly, new research has shown that CBCT gantry and

head positioning may have a positive impact on image quality in the maxillofacial region,

which in turn may improve the overall accuracy of CBCT-derived AM constructs.

In the second step of the AM workflow (Figure 1, step 2), the major challenge is

the conversion of grey scale voxel images (DICOM) into a triangulated 3D surface model

(STL). This step is necessary since all CAD software packages approved by the United

States Food and Drug Administration (FDA) require STL file formats for anatomical

modelling and reverse engineering. The DICOM to STL conversion still requires multiple

manual processing steps that are dependent on the spatial abilities and knowledge of

anatomy of the technicians or physicians. The most important step in the DICOM to STL

conversion process is image segmentation, which refers to the partitioning of images into

regions of interest (ROIs) that correspond to specific anatomical structures. A wide range

of different image segmentation methods have been developed over the past decades

that include global thresholding, edge detection, region growing, statistical shape model

(SSM)- and atlas-based methods, morphological snakes, active contouring, random forests

and artificial neural networks [29]. Of these methods, the most commonly used image

segmentation method to date in medical AM is global thresholding. It should be noted

that most software packages only offer a single, default threshold value for compact bone,

soft tissue and cartilage. In addition, these default values are often not optimized for all

types of MDCT, DECT and CBCT images and do not take into account the variations in

grey values between the different scanners. Therefore, in most cases, manual threshold

selection is necessary to acquire an optimal STL model. Threshold selection still remains,

however, a subjective task that impedes the reproducibility and reliability of medical AM

constructs. Moreover, global thresholding often causes data loss in the STL model. Such

data loss can cause voids in the STL models that can subsequently lead to inaccuracies in

3D printed implants and lead to complications during surgery [30].

The aforementioned image segmentation process has been identified as one of

the major causes of inaccuracies in medical AM constructs [31]. However, adequate

GE

NE

RA

L INTR

OD

UC

TION

12

1methods to assess the accuracy and reliability of medical AM constructs are still lacking.

More specifically, novel methods are still being sought to validate the medical imaging

and image processing steps in the medical AM workflow (Figure 1, step 1 and 2). One

method of validating imaging devices and software packages is to use phantoms that

provide ground truth measurements in clinical settings. Such phantoms are commonly

composed of materials with realistic tissue (radio)densities. However, a major drawback

with such phantoms is that they are manufactured in generic forms and sizes that do not

resemble real patients. Thus, it is difficult to extrapolate the performance of an imaging

system in phantoms to humans. Fortunately, 3D printing offers new possibilities to

fabricate anthropomorphic phantoms that can be used to assess the accuracy of different

CT scanners and software packages.

Currently, the number of patients treated with medical AM constructs is rapidly increasing.

Accordingly, the number of CT and CBCT scans performed worldwide have subsequently

increased. In this context, it should be noted that all X-ray based imaging modalities

induce harmful ionizing radiation to the patient [32]. According to the ALARA (as low as

reasonably achievable) principle, the cumulative lifetime radiation dose of patients should

be minimised by reducing the overall number of examinations and by lowering the dose

resulting from each individual examination [33]. Therefore, every CT scan that is used for

medical AM needs foremost to be suitable for diagnostic purposes. Furthermore, novel

ionizing radiation-free MRI modalities are being developed to image the musculoskeletal

system [34]. MRI provides a noninvasive visualization of water microenvironments and

other sources of protons in the human body, which provides excellent imaging possibilities

for soft tissues. Cortical bone, however, has very short relaxation times [35], leading to fast

decay of the MRI signal. As a result, cortical bone commonly appears as a signal void on

MRI images obtained with routine clinical pulse sequences such as spin echo and gradient

echo. With ultra-short echo time (UTE) sequences, an MR signal can be detected directly

from bone [36], [37]. Therefore, UTE MRI sequences may be suitable to generate images

for medical AM purposes. However, the accuracy of UTE-derived STL models still needs to

be thoroughly evaluated. In addition, MRI remains an expensive imaging modality when

compared to CT.

In conclusion, AM offers possibilities for fabricating different types of patient-specific

constructs such as anatomical models, surgical guides and implants. However, despite

recent advances and promising case studies involving medical AM, notable scientific,

technological and regulatory challenges remain.

GENERAL AIM AND OUTLINE OF THE THESISThe aim of this thesis is to provide an insight into the current scientific and technological

challenges faced in medical additive manufacturing with respect to imaging and image

processing. More specifically, the different parameters that influence the accuracy of

patient-specific AM constructs will be identified.

GE

NE

RA

L INTR

OD

UC

TION

13

1REFERENCES1. Hull C W (1986) Apparatus for production

of three-dimensional objects by stereolithography. Google Patents, 1986

2. Campbell I, Bourell D, and Gibson I (2012) Additive manufacturing: rapid prototyping comes of age. Rapid Prototyp. J. 18:255–258

3. Gibson I, Rosen D, and Stucker B (2014) Additive manufacturing technologies - 3D Printing, Rapid Prototyping, and Direct Digital Manufacturing. Springer, New York

4. Wohlers Associates (2017) 3D Printing and Additive Manufacturing State of the Industry, Annual Worldwide Progress Report. Fort Collins, CO

5. Yan Y, Li S, Zhang R, Lin F, Wu R, Lu Q, Xiong Z, and Wang X (2009) Rapid Prototyping and Manufacturing Technology: Principle, Representative Technics, Applications, and Development Trends. Tsinghua Sci. Technol. 14, Supple:1–12

6. Ventola C L (2014) Medical Applications for 3D Printing : Current and Projected Uses. Pharm. Ther. 39:704–711

7. D’Urso P S, Barker T M, Earwaker W J, Bruce L J, Atkinson R L, Lanigan M W, Arvier J F, and Effeney D J (1999) Stereolithographic biomodelling in cranio-maxillofacial surgery: a prospective trial. J. Cranio-Maxillofacial Surg. 27:30–37

8. Paiva W S, Amorim R, Bezerra D A F, and Masini M (2007) Aplication of the stereolithography technique in complex spine surgery. Arq. Neuropsiquiatr. 65:443–445

9. Dawood A, Marti B M, Sauret-Jackson V, and Darwood A (2015) 3D printing in dentistry. Br Dent J 219:521–529

10. Abou-El Fetouh A, Barakat A, and Abdel-Ghany K (2011) Computer-guided rapid-prototyped templates for segmental

mandibular osteotomies: a preliminary report. Int. J. Med. Robot. 7:187–192

11. Cohen A, Laviv A, Berman P, Nashef R, and Abu-Tair J (2009) Mandibular reconstruction using stereolithographic 3-dimensional printing modeling technology. Oral Surg. Oral Med. Oral Pathol. Oral Radiol. Endod. 108:661–6

12. Fan X, Zhou H, Lin M, Fu Y, and Li J (2007) Late reconstruction of the complex orbital fractures with computer-aided design and computer-aided manufacturing technique. J. Craniofac. Surg. 18:665–673

13. Wolff J, Sándor G K B, Pyysalo M, Miettinen A, Koivumäki A-V, and Kainulainen V T (2013) Late reconstruction of orbital and naso-orbital deformities. Oral Maxillofac. Surg. Clin. North Am. 25:683–95

14. Stoor P, Suomalainen A, Lindqvist C, Mesimäki K, Danielsson D, Westermark A, and Kontio R K (2014) Rapid prototyped patient specific guiding implants for orbital wall defects. J. Cranio-Maxillofacial Surg. 42:1644–1649

15. Rotaru H, Schumacher R, Kim S-G, and Dinu C (2015) Selective laser melted titanium implants: a new technique for the reconstruction of extensive zygomatic complex defects. Maxillofac. Plast. Reconstr. Surg. 37:1

16. Jardini A L, Larosa M A, Maciel Filho R, Zavaglia C A D C, Bernardes L F, Lambert C S, Calderoni D R, and Kharmandayan P (2014) Cranial reconstruction: 3D biomodel and custom-built implant created using additive manufacturing. J. Cranio-Maxillofacial Surg. 42:1877–1884

17. Albano T, Grabowsky M B, Rodriguez L, Lavelle W, Uhl R, Ledet E, and Sanders G (2011) Designing patient-specific orthopaedic mesh implants to treat high-energy tibial fractures. Bioengineering Conference (NEBEC), 2011 IEEE 37th Annual Northeast

GE

NE

RA

L INTR

OD

UC

TION

14

118. Visscher D O, Farré-Guasch E, Helder M

N, Gibbs S, Forouzanfar T, van Zuijlen P P, and Wolff J (2016) Advances in Bioprinting Technologies for Craniofacial Reconstruction. Trends Biotechnol. 34:700-710

19. Bakhos D, Velut S, Robier A, Al Zahrani M, and Lescanne E (2010) Three-Dimensional Modeling of the Temporal Bone for Surgical Training. Otol. Neurotol. 31:328–334

20. Bruyère F, Leroux C, Brunereau L, and Lermusiaux P (2008) Rapid prototyping model for percutaneous nephrolithotomy training. J. Endourol. 22:91–6

21. McMenamin P G, Quayle M R, McHenry C R, and Adams J W (2014) The production of anatomical teaching resources using three-dimensional (3D) printing technology. Anat. Sci. Educ. 7:479–486

22. Bibb R, Eggbeer D, and Paterson A (2015) Medical modelling: The application of advanced design and rapid prototyping techniques in medicine. Woodhead Publishing, Cambridge, UK

23. Lorensen W E and Cline H E (1987) Marching cubes: A high resolution 3D surface construction algorithm. ACM siggraph computer graphics. 21:163–169

24. EU (2017) Revisions of medical device directives - European Commission (2017). Available: https://ec.europa.eu/growth/sectors/medical-devices/regulatory- f ramework/revis ion_en. [Accessed: 30-May-2017]

25. Liang X, Lambrichts I, Sun Y, Denis K, Hassan B, Li L, Pauwels R, and Jacobs R (2010) A comparative evaluation of Cone Beam Computed Tomography (CBCT) and Multi-Slice CT (MSCT). Part II: On 3D model accuracy. Eur. J. Radiol. 75:270–274

26. Mallepree T and Bergers D (2009) Accuracy of medical RP models. Rapid Prototyp. J. 15:325–332

27. Pauwels R, Beinsberger J, Collaert B, Theodorakou C, Rogers J, Walker A, Cockmartin L, Bosmans H, Jacobs

R, Bogaerts R, and Horner K (2012) Effective dose range for dental cone beam computed tomography scanners. Eur. J. Radiol. 81:267–271

28. Fahrig R, Dixon R, Payne T, Morin R L, Ganguly A, and Strobel N (2006) Dose and image quality for a cone-beam C-arm CT system. Med. Phys. 33:4541–4550

29. Sharma N and Aggarwal L M (2010) Automated medical image segmentation techniques. J. Med. Phys. 35:3–14

30. Stoor P, Suomalainen A, Lindqvist C, Mesimaki K, Danielsson D, Westermark A, and Kontio R K (2014) Rapid prototyped patient specific implants for reconstruction of orbital wall defects. J. Craniomaxillofac. Surg. 42:1644–1649

31. Smith E J, Anstey J A, Venne G, and Ellis R E (2013) Using additive manufacturing in accuracy evaluation of reconstructions from computed tomography. Proc Inst Mech Eng H 227:551–559

32. Koivisto J (2015) The use of MOSFET dosimeters and anthropomorphic phantoms in low dose dental CBCT applications. Helsinki, Finland: University of Helsinki

33. Implications of Commission Recommendations that Doses be kept as Low as Readily Achievable (1973) ICRP publication 22, Oxford, UK

34. Delso G, Carl M, Wiesinger F, Sacolick L, Porto M, Hüllner M, Boss A, and Veit-Haibach P (2014) Anatomic Evaluation of 3-Dimensional Ultrashort-Echo-Time Bone Maps for PET/MR Attenuation Correction. J. Nucl. Med. 55:780–785

35. Horch R A, Nyman J S, Gochberg D F, Dortch R D, and Does M D (2010) Characterization of 1H NMR signal in human cortical bone for magnetic resonance imaging. Magn. Reson. Med. 64:680–687

36. Keereman V, Fierens Y, Broux T, De Deene Y, Lonneux M, and Vandenberghe S (2010) MRI-based attenuation correction

GE

NE

RA

L INTR

OD

UC

TION

15

1for PET/MRI using ultrashort echo time sequences. J. Nucl. Med. 51:812–8

37. Aitken A P, Giese D, Tsoumpas C, Schleyer P, Kozerke S, Prieto C, and

Schaeffter T (2014) Improved UTE-based

attenuation correction for cranial

PET-MR using dynamic magnetic field

monitoring. Med. Phys. 41:12302

2 A NOVEL METHOD OF ORBITAL FLOOR RECONSTRUCTION

USING VIRTUAL PLANNING, 3-DIMENSIONAL PRINTING, AND AUTOLOGOUS BONE

Maarten Vehmeijer, Maureen van Eijnatten,

Niels Liberton, Jan Wolff

Journal of Oral and Maxillofacial Surgery (2016). 74 (8): 1608-1612doi: 10.1016/j.joms.2016.03.044

A N

OV

EL M

ETH

OD

OF O

RB

ITAL FLO

OR

RE

CO

NSTR

UC

TION

18

2

ABSTRACTFractures of the orbital floor are often a result of traffic accidents or interpersonal violence.

To date, numerous materials and methods have been used to reconstruct the orbital

floor. However, simple and cost-effective 3-dimensional (3D) printing technologies for

the treatment of orbital floor fractures are still sought. This study describes a simple,

precise, cost-effective method of treating orbital fractures using 3D printing technologies

in combination with autologous bone. Enophthalmos and diplopia developed in a 64-

year-old female patient with an orbital floor fracture. A virtual 3D model of the fracture

site was generated from computed tomography images of the patient. The fracture was

virtually closed using spline interpolation. Furthermore, a virtual individualized mould of

the defect site was created, which was manufactured using an inkjet printer. The tangible

mould was subsequently used during surgery to sculpture an individualized autologous

orbital floor implant. Virtual reconstruction of the orbital floor and the resulting mould

enhanced the overall accuracy and efficiency of the surgical procedure. The sculptured

autologous orbital floor implant showed an excellent fit in vivo. The combination of virtual

planning and 3D printing offers an accurate and cost-effective treatment method for orbital

floor fractures.

A N

OV

EL M

ETH

OD

OF O

RB

ITAL FLO

OR

RE

CO

NSTR

UC

TION

19

2

INTRODUCTIONBlow-out fractures are often a result of motor vehicle accidents, interpersonal violence, or

sports injuries. Such events can induce a blunt impact to the orbital area, which increases

the intraorbital pressure and subsequently leads to a fracture of one or more orbital walls.

This often results in a prolapse of the orbital content and, hence, displacement of the eye

globe into the maxillary sinus cavity, which can lead to a enophthalmos (infra-position of

the globe) and diplopia (double vision) of the eye.

To date, a wide range of different reconstruction materials have been used for

the treatment of orbital floor fractures. Nevertheless, over the past few decades, autogenous

bone grafts have been considered the gold standard [1]. However, a major drawback of

autologous bone compared with prefabricated titanium meshes is that autologous bone

is cumbersome to sculpture and can easily break if bent beyond its capacity. Furthermore,

bone grafts can in some cases resorb unpredictably. However, orbital graft resorption can

be minimized by proper fixation of the graft to the orbital floor [2].

Recent advancements in head imaging and 3-dimensional (3D) printing have opened

a wide range of new treatment possibilities [3]. Such technologies allow the creation of

patient-specific medical constructs. Currently, medical 3D printing is most commonly

used to manufacture medical models, drill guides, saw guides and individualized titanium

reconstruction plates [4], [5].

To date, biocompatible 3D printable materials are still being sought. Therefore, we

suggest combining currently available 3D printing technologies with autologous bone

grafts for the reconstruction of the orbital floor. The aim of this study was to develop

a simple and cost-effective method to produce individualized orbital implants using

computed tomography (CT) imaging and virtual planning, 3D printed moulds, and

autologous bone grafts.

MATERIALS AND METHODSA swollen and painful left eye developed in a 64-year-old female patient after a bicycle

accident. Clinical examination showed a left-sided periorbital hematoma, subconjuntival

ecchymosis, mild enophtalmus, and diplopia. Furthermore, the patient’s left eye movement

was restricted in the vertical plane. CT images showed a blowout fracture of the left

orbit (Figure 1).

The CT images were acquired using a GE Discovery CT750 HD Multi-Slice CT

scanner (GE Healthcare, Waukesha, WI). The following imaging parameters were used:

tube voltage, 120 kV; tube current, 300 mA; and reconstruction kernel, soft. The slices

were 0.6 mm thick, with 0.312-mm gaps between them. The resulting Digital Imaging

and Communications in Medicine (DICOM) files were imported into OsiriX MD imaging

software (Pixmeo, Geneva, Switzerland), and a virtual 3D model of the patient’s skull was

created using a threshold value of 120 Hounsfield units.

A N

OV

EL M

ETH

OD

OF O

RB

ITAL FLO

OR

RE

CO

NSTR

UC

TION

20

2

The virtual 3D model was used to manufacture a 3D printed mould of the virtually

reconstructed orbital floor. The virtual reconstruction was performed as follows:

The boundary of the orbital floor defect was delineated in the virtual 3D model using

3-matic Medical 9.0 software (Materialise, Leuven, Belgium). In a second step, 7 spline

interpolation curves were virtually drawn through the defect site to reconstruct the missing

orbital floor bone (Figure 2). In a third step, a rectangular box (60 x 60 x 30 mm) was

created around the reconstructed virtual 3D model of the orbital floor. The rectangular

box was necessary to 3D print a stable tangible mould of the orbital floor (Figure 3),

allowing the surgeon to apply force whilst sculpturing the autologous bone graft.

The aforementioned virtual reconstruction of the orbital floor took approximately 45

minutes. After successful virtual reconstruction, one tangible mould was printed using

a ZCorp inkjet printer (3D Systems, Rock Hill, SC). The total printing time required to

fabricate the tangible mould was 84 minutes.

Figure 1. Computed tomography image (coronal view) showing a blowout fracture of the left orbit with protruding soft tissue into the maxillary sinus (arrow).

A N

OV

EL M

ETH

OD

OF O

RB

ITAL FLO

OR

RE

CO

NSTR

UC

TION

21

2

During surgery, bone was harvested from the anterior iliac crest. The resulting bone

graft was thinned down to 1 mm in the outer area and to 2 to 3 mm in the central area.

The previously printed mould of the orbital floor was wrapped in a sterile plastic bag and

used as a template to mould and sculpture the graft into an individualized autologous

orbital floor implant (Figure 3).

The individualized implant was meticulously placed into the orbital floor and

fixated with two 3-hole miniplates (1.0 mm) and four 3-mm centre drive screws (KLS

Martin, Tuttlingen, Germany) to the inferior orbital rim. The surgical procedure took 40

minutes. Postoperative CT images revealed a harmonious reconstruction of the orbital

floor (Figure 4).

Figure 2. Virtual 3-dimensional model of the patient’s skull (grey), including the virtual mould of the orbital floor (A), the virtually reconstructed bone defect (B), and the spline interpolation curves (C).

Figure 3. Autologous bone graft (ie, implant) sculptured and moulded on a printed mould (A) with a reconstructed orbital floor (B).

A N

OV

EL M

ETH

OD

OF O

RB

ITAL FLO

OR

RE

CO

NSTR

UC

TION

22

2

RESULTS AND DISCUSSIONA wide variety of materials have been described for the treatment of orbital floor fractures,

ranging from polymers to metals [6]. In the past decade, individualized, 3D printed

titanium orbital implants have been used in maxillofacial surgery. However, such implants

are difficult to fabricate, are expensive, and sometimes lack dimensional precision [7].

Furthermore, they are difficult to adapt or sculpture during surgery because of their

mechanical properties. Medical titanium alloys typically have an elastic modulus (E-

modulus or Young’s modulus) of 120 GPa [8], whilst the E-modulus of orbital bone ranges

between 1.26 and 4.55 GPa [9]. Such differences can cause irritation and resorption of

the bone underneath the titanium implant [10]. The major advantage of the moulded

autologous bone implant used in this study is its biocompatibility, which subsequently

causes less irritation to surrounding tissues.

Historically, mirroring of the contralateral healthy orbit has been commonly used in

virtual planning and subsequent printing of orbital implants. However, this technique

has certain drawbacks. The major drawback is that the human face is often asymmetric

and therefore geometric variability between the two orbits is possible. Therefore,

Figure 4. Postoperative computed tomography image of the autologous orbital floor implant in situ.

A N

OV

EL M

ETH

OD

OF O

RB

ITAL FLO

OR

RE

CO

NSTR

UC

TION

23

2

we recommend using spline interpolation that offers the possibility of closing orbital

defects using the surrounding healthy bony structures as a reference. However, creating

an accurate spline-based mould requires technical and computational knowledge that

may be challenging for most clinicians. This problem could be solved using large image

databases combined with machine learning and pattern recognition algorithms that would

automatically reconstruct complex anatomic geometries in trauma patients.

The orbital floor mould used in this study was printed in 84 minutes using a high-

resolution 3D printer. During fabrication, the printer deposits a 0.1-mm-thick layer of powder

over which the printing head moves and deposits a binder that solidifies the powder.

The thin-layer increments subsequently offer very precise 3D printed constructs.

However, designing the very precise virtual model required for the fabrication of a mould

is slightly more challenging since it is dictated by the correctness of the conversion of CT

images into virtual 3D models (Figure 2). The accuracy of a virtual 3D model derived from

CT images currently ranges between 0.4 and 0.7 mm [11]. Nevertheless, the autologous

orbital floor implant used in this case report demonstrated an excellent fit in-situ and

subsequently reduced the operation time by approximately 40% (Figure 4).

The total manufacturing costs of the orbital floor mould were low (€85) compared with

other treatment options. To date, 3D printed titanium orbital implants have cost between

€3,000 and €5,000 euros. The costs of prefabricated titanium orbital plates or meshes

range between €250 euros (KLS Martin) and €600 and €800 (Synthes, West Chester, PA).

The treatment of bony defects with autologous bone still remains a challenge since

the amount of donor bone available is anatomically limited. A major drawback of

autogenous bone is, of course, the need for harvesting, which is accompanied by donor-

site morbidity. In addition, the sculpturing of autologous bone to fit the recipient site can

be time-consuming and manually cumbersome and subsequently result in inconsistent

cosmetic outcomes. However, this problem can be minimized using the treatment method

described in this study.

In conclusion, virtual preoperative planning combined with 3D printing offers

a geometrically predictable and low-cost method of creating autologous individualized

orbital floor implants.

A N

OV

EL M

ETH

OD

OF O

RB

ITAL FLO

OR

RE

CO

NSTR

UC

TION

24

2

REFERENCES1. Kontio R K, Laine P, Salo A, Paukku P,

Lindqvist C, and Suuronen R (2006) Reconstruction of internal orbital wall fracture with iliac crest free bone graft: clinical, computed tomography, and magnetic resonance imaging follow-up study. Plast. Reconstr. Surg. 118:1365–1374

2. Glassman R D, Manson P N, Vanderkolk C A, Iliff N T, Yaremchuk M J, Petty P, Defresne C R, and Markowitz B L (1990) Rigid fixation of internal orbital fractures. Plast. Reconstr. Surg. 86:1101–1103

3. Wolff J, Sándor G K B, Pyysalo M, Miettinen A, Koivumäki A-V, and Kainulainen V T (2013) Late reconstruction of orbital and naso-orbital deformities. Oral Maxillofac. Surg. Clin. North Am. 25:683–95

4. Kang S-H, Lee J-W, Lim S-H, Kim Y-H, and Kim M-K (2014) Verification of the usability of a navigation method in dental implant surgery: in vitro comparison with the stereolithographic surgical guide template method. J. Craniomaxillofac. Surg. 42:1530–5

5. Mazzoni S, Bianchi A, Schiariti G, Badiali G, and Marchetti C (2015) Computer-aided design and computer-aided manufacturing cutting guides and customized titanium plates are useful in upper maxilla waferless repositioning. J. Oral Maxillofac. Surg. 73:701–707

6. Avashia Y J, Sastry A, Fan K L, Mir H S, and Thaller S R (2012) Materials used for

reconstruction after orbital floor fracture. J. Craniofac. Surg. 23:1991–1997

7. Stoor P, Suomalainen A, Lindqvist C, Mesimaki K, Danielsson D, Westermark A, and Kontio R K (2014) Rapid prototyped patient specific implants for reconstruction of orbital wall defects. J. Craniomaxillofac. Surg. 42:1644–1649

8. Schuh A, Bigoney J, Hönle W, Zeiler G, Holzwarth U, and Forst R (2007) Second generation (low modulus) titanium alloys in total hip arthroplasty. Materwiss. Werksttech. 38:1003–1007

9. Ong H (2012) Biomechanical properties of the human orbital floor for selection of reconstruction materials using a mathematical engineering model [master’s thesis]. Groningen, The Netherlands, University of Goningen, 2012. Available at: http://irs.ub.rug.nl/dbi/53f59f60d8234

10. Al-Sukhun J, Penttila H, and Ashammakhi N (2012) Orbital stress analysis, Part IV: Use of a “stiffness-graded” biodegradable implants to repair orbital blow-out fracture. J. Craniofac. Surg. 23:126–130

11. van Eijnatten M, Rijkhorst E-J, Hofman M, Forouzanfar T, and Wolff J (2016) The accuracy of ultrashort echo time MRI sequences for medical additive manufacturing. Dentomaxillofacial Radiol. 45:20150424

3 INFLUENCE OF CT PARAMETERS ON STL MODEL ACCURACY

Maureen van Eijnatten, Ferco Berger, Pim de Graaf,

Juha Koivisto, Tymour Forouzanfar, Jan Wolff

Rapid Prototyping Journal (2017). 24 (4): 679-685doi: 10.1108/RPJ-07-2015-0092

INFLU

EN

CE

OF C

T PAR

AM

ETE

RS O

N STL M

OD

EL A

CC

UR

AC

Y

28

3

ABSTRACTPurpose Additive manufactured (AM) skull models are increasingly used to plan complex

surgical cases and design custom implants. The accuracy of such constructs depends

on the standard tessellation language (STL) model, which is commonly obtained from

computed tomography (CT) data. The aims of this study were to assess the image quality

and the accuracy of STL models acquired using different CT scanners and acquisition

parameters.

Design/methodology/approach Images of three dry human skulls were acquired using two multi-detector row computed

tomography (MDCT) scanners, a dual energy computed tomography (DECT) scanner and

one cone beam computed tomography (CBCT) scanner. Different scanning protocols

were used on each scanner. All images were ranked according to their image quality and

converted into STL models. The STL models were compared to gold standard models.

Findings Image quality differed between the MDCT, DECT and CBCT scanners. Images acquired

using low-dose MDCT protocols were preferred over images acquired using routine

protocols. All CT-based STL models demonstrated non-uniform geometrical deviations of

up to +0.9 mm. The largest deviations were observed in CBCT-derived STL models.

Practical implications While patient-specific AM constructs can be fabricated with great accuracy using AM

technologies, their design is more challenging because it is dictated by the correctness

of the STL model. Inaccurate STL models can lead to ill-fitting implants that can cause

complications after surgery.

Originality/value This paper suggests that CT imaging technologies and their acquisition parameters affect

the accuracy of medical AM constructs.

INFLU

EN

CE

OF C

T PAR

AM

ETE

RS O

N STL M

OD

EL A

CC

UR

AC

Y

29

3

INTRODUCTIONAdditive manufacturing (AM), often referred to as three-dimensional (3D) printing, is

becoming increasingly popular in the field of medicine. Recent developments in high-

resolution imaging techniques and more affordable 3D printers have led to an increased

use of different AM technologies in clinical settings [1]. AM medical models are becoming

indispensable in medical education and communication [2], [3], treatment planning, and

surgical guidance [4], and have been shown to improve surgical outcomes and reduce

operating times [5]. Furthermore, AM drill and saw guides in combination with patient-

specific constructs such as implants or reconstruction plates are being more readily used

in maxillofacial surgery [6].

Medical AM differs from industrial AM in that it requires patient-specific 3D medical

images. To date, such images are commonly acquired using computed tomography (CT)

technology. The resulting CT datasets are archived as Digital Imaging and Communications

in Medicine (DICOM) files that are subsequently converted into standard tessellation

language (STL) models. Such models are necessary to design medical AM constructs using

commercially available computer-aided design (CAD) software packages.

A wide range of different CT technologies are currently available on the market for

diagnostic purposes. To date, multi-detector row computed tomography (MDCT) scanners

are most commonly used in hospital environments. These scanners have multiple rows

of detectors, hence short scanning times, and allow high-resolution imaging of bone

[7]. In addition, dual energy computed tomography (DECT) scanners have recently been

developed and offer enhanced tissue characterization by using two X-ray beams with

different photon energy spectra [8]. However, the scanning protocols used with these

advanced imaging technologies exhibit a wide variability in terms of image acquisition

and exposure parameters. These parameters include the scan mode (e.g. axial or helical),

tube potential (kV), exposure (mAs), rotation time, X-ray beam filter type and size, beam

collimation, detector configuration, helical pitch (or table feed per gantry rotation), and

field of view (FOV).

Cone beam computed tomography (CBCT) scanners are commonly used in private

clinics due to their small size and low costs. CBCT technology uses a cone shape X-ray

beam and a single flat panel detector to obtain 3D information of the head of a patient

by data acquisition through a single 200°-360° gantry rotation [9]. Cone beam technology

induces lower radiation doses compared with MDCT [10], [11] and has recently become

more affordable. This has led to an increase in popularity amongst maxillofacial radiologists

and surgeons [12]. Comparable to MDCT, modern CBCT scanners offer a wide range of

image acquisition parameters that include scan time, tube potential (kV), exposure (mAs),

rotation time, gantry rotation, FOV, and various metal artifact reduction algorithms.

The aforementioned CT technologies and resulting datasets are currently used for

diagnostic purposes in clinical settings. Such diagnostic datasets are also pro tempore

used to design medical constructs. However, it must be noted that imaging protocols

INFLU

EN

CE

OF C

T PAR

AM

ETE

RS O

N STL M

OD

EL A

CC

UR

AC

Y

30

3

required for diagnostic purposes may differ markedly from those required by technicians

to design patient-specific AM constructs [13]. Nevertheless, all clinicians are bound to

the ALARA (as low as reasonably achievable) radiation safety principle. Therefore, all CT

scans required for patient-specific constructs need to be foremost diagnostable. The aims

of this study were to assess the image quality and the accuracy of STL models acquired

using different MDCT and CBCT scanners and acquisition parameters.

MATERIALS AND METHODSThree formalin-fixed human cadaver skulls of succumbed patients with intact bony structures

were obtained from the Department of Anatomy of the VU University Medical Center,

Amsterdam. All soft tissues were meticulously removed. The skulls were subsequently

scanned using a GOM ATOS™ optical 3D scanner (GOM, Braunschweig, Germany). This

device uses non-contact, blue-light technology to record a series of images of an object

from different angles, and combines these images into a triangular mesh representing

the surface of the object. The GOM ATOS™ optical 3D scanner offers a point accuracy

of 0.05 mm and was therefore used to obtain a “gold standard” STL model of each

skull (Figure 1).

Figure 1. Outline of the study.

INFLU

EN

CE

OF C

T PAR

AM

ETE

RS O

N STL M

OD

EL A

CC

UR

AC

Y

31

3

CT image acquisitionAll three dry human skulls were subsequently scanned using two different MDCT scanners

and a CBCT scanner (Figure 1): a Siemens Somatom Sensation 64-slice MDCT scanner

(Siemens Medical Solutions, Erlangen, Germany), a GE Discovery CT750 HD 64-slice

MDCT scanner (GE Healthcare, Little Chalfont, Buckinghamshire, UK), and a Vatech PaX

Zenith 3D CBCT scanner (Vatech, Fort Lee, USA). In addition, the GE scanner offered

a DECT image acquisition possibility using rapid switching X-ray beams with 140 kVp and

80 kVp voltage. All acquisition parameters are presented in Table 1.

Different scanning protocols were used on the aforementioned MDCT and CBCT

scanners. Initially, routine scanning protocols recommended by the device manufacturers

were used. In a second step, minimum and maximum tube potential (kV) and tube current

(mA) were used (Table 2) in order to assess the influence of these parameters on image

quality and accuracy of the STL models. Furthermore, the radiation output was calculated

using the dose-length product (DLP) for all MDCT protocols, and the dose-area product

(DAP) for all CBCT protocols.

All MDCT and DECT images were reconstructed using sharp reconstruction kernels

(Siemens: H70h, GE: BONEPLUS for routine mode and DETAIL for dual energy mode).

These kernels were selected after assessing all the different kernels available on the Siemens

and GE scanners.

Image quality analysisFollowing image acquisition and reconstruction, two radiologists with specific expertise

in maxillofacial bone imaging visually assessed and ranked all 57 DICOM datasets hence

Table 1. MDCT, DECT and CBCT exposure parameters.

Exposure parameters

Siemens Somatom Sensation 64-slice MDCT

GE Discovery CT750 HD 64-slice MDCT/DECT

Vatech PaX Zenith CBCT

Tube voltage (kVp) 80-140 80-140 50-120Tube current (mA) 150-650 150-650 5-10Rotation time (s) 1 1 15-24Q (mAs) 150-650 150-650 75-240Slice thickness (mm) 0.6 0.625 n.a.Slice increment (mm) 0.391 0.312 n.a.Pitch (mm) 0.55-1 0.5-1 n.aCTDIvol (mGy) 5.94-113 17.7-227 n.a.Voxel (mm) 0.391 x 0.391 x 0.6 0.312 x 0.312 x 0.625 0.2 x 0.2 x 0.2Voxel number 512 x 512 x 512 512 x 512 x 512 1056 x 1056 x 1056Scan angle 360º 360º 360ºPulsed No Only for dual energy YesScan FOV (cm) 20 20 Free (~24x19)

INFLU

EN

CE

OF C

T PAR

AM

ETE

RS O

N STL M

OD

EL A

CC

UR

AC

Y

32

3

images (19 per skull). The radiologists were blinded from the acquisition parameters and

from the results of one another. A constant Window/Level setting was used to evaluate all

images equally (W6000; L100). The following anatomical structures were visually assessed:

1) Laminae papyracea; 2) Infra-orbital canal; 3) Sphenoid sinus septum; and 4) Orbital apex

(Figure 2). Each anatomical structure was given scores from 1 to 4 in ascending order; 1 =

anatomic structure not identifiable due to poor image quality; 2 = structure identified but

no details assessable and insufficient image quality; 3 = anatomic structure fully assessable

in all parts and acceptable image quality; 4 = very good delineation of structure and

excellent image quality.

Because of the small variations in the image quality of the different scanning protocols

(Table 3), a second more specific ranking system was necessary. In this ranking system,

all images were given scores from best to worst according to relative image quality

(Table 4). All image assessments and hence rankings were repeated after a 6-month

interval to minimize the radiologists’ learning effect.

Statistical analysis of the image qualityThe mean values and standard deviations (SD) of both ranking systems (Table 3 and 4) were

calculated using SPSS Statistics software (version 23.0 for Windows; SPSS Inc, Chicago,

USA). Furthermore, all rankings were compared using a nonparametric Mann-Whitney U

test with a 2-tailed p value. Statistical significance was established at p ≤ 0.05.

Image processingAfter image qualtiy analysis, All MDCT, DECT, and CBCT DICOM datasets were

subsequently converted into STL models using OsiriX MD software (Pixmeo, Geneva,

Switzerland). Automatic thresholding was used with a uniform intensity value of -500

Hounsfield Units (HU).

Table 2. MDCT, DECT and CBCT scanning protocols: tube potentials (kV) and exposures (mAs).

Scanning protocolMDCT (Siemens)kV / mAs

MDCT (GE)kV / mAs

DECT (GE)kV / mAs

CBCT (Vatech)kV / mAs

1. Routine 120 / 380 120 / 220 80 & 140 / 375 115 / 1442. Routine kV / max. mA 120 / 650 120 / 650 80 & 140 / 630 115 / 2403. Min. kV / max. mA 80 / 650 80 / 650 n.a.4. Max. kV / max. mA 140 / 550 140 / 510 120 / 2405. Min. kV / min. mA 80 / 150 80 / 150 50 / 756. Max. kV / Routine mA 120 / 1447. Metal artifact reduction 115 / 1448. Mediate kV / routine mA 80 / 144

INFLU

EN

CE

OF C

T PAR

AM

ETE

RS O

N STL M

OD

EL A

CC

UR

AC

Y

33

3

Standard tessellation language deviation analysisAll STL files were imported into GOM Inspect software (GOM GmbH, Braunschweig,

Germany), and were cleared of outliers and aligned with the gold standard STL model

using a “Local Best-Fit” function. Geometrical differences between the STL models and

the gold standard were calculated using the “Surface Comparison on Actual” function and

visualized using colour maps. A detailed assessment of the orbital area was performed

using the “Select on Surface” function. All distance values were exported to Excel

software (Microsoft Office 2013, Redmond, WA, United States) and the mean and SD were

calculated. Finally, histograms and kurtosis were calculated using an excel data analysis

software plugin.

RESULTSThe image quality of the following four anatomical structures was visually assessed:

laminae papyracea; infra-orbital canal; sphenoid sinus septum and orbital apex (Figure 2).

Furthermore, the DLP of each MDCT examination and the DAP of each CBCT examination

was calculated. The results are summarized in Table 3. Scanner-specific rankings of

the different scanning protocols are presented in Table 4.

All visually examined images were subsequently converted into STL models.

The resulting STL models were compared to a gold standard skull STL model acquired

using an optical scanner. When compared with the gold standard STL models, all STL

models obtained from MDCT, DECT, and CBCT datasets demonstrated large non-

uniform deviations in the orbital wall, maxillary sinus, arcus zygomaticus, and nasal cavity.

Figure 2. Evaluated anatomical structures, 1) Laminae papyracea; 2) Infra-orbital canal; 3) Spenoid sinus septum; and 4) Orbital apex. The images were acquired using a Siemens MDCT scanner, 80 kV, 150 mA, window/level: 6000/100.

INFLU

EN

CE

OF C

T PAR

AM

ETE

RS O

N STL M

OD

EL A

CC

UR

AC

Y

34

3

Tab

le 3

. Im

age

qua

lity

sco

res

(sca

le 1

– 4

) and

do

se-l

eng

th p

rod

uct

(DLP

) / d

ose

-are

a p

rod

uct

(DA

P) o

f all

MD

CT,

DEC

T, a

nd C

BC

T d

atas

ets a

.

Scan

ning

pro

toco

l

MD

CT

(Sie

men

s)M

DC

T (G

E)

DE

CT

(GE

)C

BC

T (V

atec

h)

Mea

n ±

SD

pD

LP

(mG

y-cm

)M

ean

±

SDp

DLP

(m

Gy-

cm)

Mea

n ±

SD

pD

LP

(mG

y-cm

)M

ean

±

SDD

AP

(mG

y-cm

2 )

1. R

out

ine

3 ±

0.3

– 11

323

± 0

.5–

1632

2 ±

0.4

–14

942

± 0

.6–

25.0

8

2. R

out

ine

kV /

max

. mA

3 ±

0.3

118

653

± 0

.6.7

8546

762

± 0

.41

1614

1 ±

0.5

.029

*41

.51

3. M

in. k

V /

max

. mA

2 ±

0.6

.061

544

3 ±

0.8

116

764.

Max

. kV

/ m

ax. m

A3

± 0

.0.3

1723

903

± 0

.8.4

4051

111

± 0

.3.0

01*

43.3

65.

Min

. kV

/ m

in. m

A3

± 0

.31

131

3 ±

0.7

.797

398

1 ±

0.4

.004

*1.

646.

Max

. kV

/ R

out

ine

mA

1 ±

0.3

.001

*25

.20

7. M

etal

art

ifact

red

uctio

n2

± 0

.8.7

3225

.08

8. M

edia

te k

V /

ro

utin

e m

A2

± 0

.8.2

1111

.82

a p v

alue

s in

dic

ate

a co

mp

aris

on

with

the

ro

utin

e p

roto

col (

* =

p <

.05)

. All

p v

alue

s fo

r M

DC

T an

d D

EC

T sc

anni

ng p

roto

cols

are

no

nsig

nific

ant.

Tab

le 4

. Sc

anne

r-sp

ecifi

c ra

nkin

gs

of a

ll M

DC

T, D

ECT,

and

CB

CT

dat

aset

s a .

Scan

ning

pro

toco

l

MD

CT

(Sie

men

s)M

DC

T (G

E)

DE

CT

(GE

)C

BC

T (V

atec

h)

Mea

n ±

SD

pM

ean

± S

Dp

Mea

n ±

SD

pM

ean

± S

Dp

1. R

out

ine

3.2

± 1

.3–

2.9

± 1

.1–

1.9

± 0

.2–

6.2

± 0

.9–

2. R

out

ine

kV /

max

. mA

3.1

± 1

.1.2

842.

5 ±

1.3

.023

*1.

1 ±

0.2

.000

*3.

1 ±

1.3

.000

*3.

Min

. kV

/ m

ax. m

A1.

6 ±

1.0

.000

*3.

2 ±

1.2

.154

4. M

ax. k

V /

max

. mA

3.5

± 1

.2.1

791.

9 ±

1.3

.000

*1.

3 ±

0.6

.000

*5.

Min

. kV

/ m

in. m

A3.

7 ±

1.5

.074

4.3

± 1

.1.0

00*

3.3

± 1

.1.0

00*

6. M

ax. k

V /

Ro

utin

e m

A3.

3 ±

1.7

.000

*7.

Met

al a

rtifa

ct r

educ

tion

6.1

± 1

.0.7

798.

Med

iate

kV

/ r

out

ine

mA

4.8

± 1

.1.0

00*

a p v

alue

s in

dic

ate

a co

mp

aris

on

with

the

ro

utin

e p

roto

col (

* =

p <

.05)

.

INFLU

EN

CE

OF C

T PAR

AM

ETE

RS O

N STL M

OD

EL A

CC

UR

AC

Y

35

3

The calculated deviations of one skull (no. 2) are presented in Figure 3. The other two

skulls (no. 1, 3) presented similar deviations.

The STL models of skull 1, 2, and 3 acquired using the Siemens MDCT scanner showed

deviations of 0.58 mm, 0.49 mm, and 0.77 mm and SDs of 0.91, 0.81, 0.99, respectively

(Figure 3A). The GE MDCT STL deviations ranged from 0.56 mm, 0.45 mm, and 0.74 mm

with SDs of 0.88, 0.76, and 0.96, respectively (Figure 3B). DECT STL models resulted

in deviations of 0.44 mm, 0.45 mm, and 0.62 mm with SDs of 0.73, 0.71, and 0.84,

respectively (Figure 3C). Furthermore, DECT STL models showed less detail in the orbital

and maxillary sinus area when compared with the MDCT models. Finally, the CBCT scanner

presented deviations of 0.63 mm, 0.63 mm, and 0.80 mm with SDs of 0.88, 0.88, and 0.95,

respectively (Figure 3D).

The deviations between the gold standard STL models and STL models acquired from

all three skulls using four different CT scanners generally ranged between -0.9 mm and

Figure 3. Geometric deviations mapped on the surface of the STL models acquired using a routine scanning protocol on four different CT scanners. (a) MDCT (Siemens) - 120 kV/380mAs; (b) MDCT (GE) - 120 kV/380 mAs; (c) DECT (GE) - 80&140 kV/375 mAs; (d) CBCT (Vatech) - 115 kV/144 mAs

INFLU

EN

CE

OF C

T PAR

AM

ETE

RS O

N STL M

OD

EL A

CC

UR

AC

Y

36

3

Figure 4. Frequency of the geometric deviations (in mm) of the STL models of all three skulls when compared with the gold standard.

+0.8 mm and are summarized in a histogram (Figure 4). The MDCT STL models showed

the smallest deviations, but the models were on average larger than the gold standard

STL model. The kurtosis – the steepness of a curve compared to a normal curve – of

the deviations in the MDCT STL models was 1.7 (Siemens) and 2.7 (GE), 0.08 in the DECT

STL models and -0.9 in the CBCT STL models.

Detailed assessments of the orbital area in two Siemens MDCT STL models of skull

2 are presented in Figure 5. When compared with the routine scanning protocol (A),

the min kV/min mA protocol (B) resulted in better anatomical details. The mean deviation

in the orbits was 0.25 ± 0.41 mm (routine protocol) and 0.29 ± 0.44 mm (min mA/min kV

protocol). In all MDCT STL models, the orbits were larger than the gold standard in 80 per

cent of the surface area.

Figure 5. An example of the deviations in the STL models attained using standard and min kV/min mA scanning protocols on the Siemens MDCT scanner.

INFLU

EN

CE

OF C

T PAR

AM

ETE

RS O

N STL M

OD

EL A

CC

UR

AC

Y

37

3

DISCUSSIONIn the present study, the image quality and the accuracy of skull-derived STL models

acquired using different MDCT and CBCT scanners and acquisition parameters

were assessed.

The image quality differed markedly between the MDCT, DECT and CBCT scanners

(Table 3). The MDCT scanners offered the best image quality. Furthermore, the images

acquired using min kV/min mA and hence low-dose MDCT scanning protocols were

preferred over the images acquired using routine protocols (Table 4). Moreover, compared

to the routine protocols, the low-dose protocols showed a significant reduction in the DLP

of 88 and 76 per cent on the Siemens and GE scanner, respectively (Table 3). These findings

have large clinical implications since the use of low-dose scanning protocols can markedly

reduce the effective dose in patients who require patient-specific AM constructs without

jeopardizing the overall diagnostic image quality.

After image quality assessment, the acquired DICOM datasets were converted into STL

models and geometrically compared to the corresponding gold standard skull models.

In this context, the MDCT scanners offered the most accurate STL models (Figure 3) with

absolute mean deviations ranging between 0.45 mm and 0.77 mm when compared with

the optical scan “gold standard” STL model. These results are in good agreement with

the findings of Thorhauer et al. (2010), who found mean deviations of 0.512 mm between

MDCT STL models and optical scan models of the knee [14]. Other studies reported mean

deviations of 0.13 mm [15] and 0.45 mm [16] on long cadaver bones and 0.137 mm on dry

human mandibles [17]. However, these studies only included large bony structures.

The STL models acquired using DECT scanning protocols demonstrated greater loss of

bony details when compared to the MDCT protocols, especially in the thin bony structures

of the orbit (Figure 3). Deviations ranging between 0.44 mm and 0.62 mm were found in

the DECT STL models. These findings could be due to the rapid switching between photon

energy spectra required for dual energy imaging on the GE scanner. Another explanation

could be the lack of a bone reconstruction kernel for dual energy image reconstruction.

The largest geometrical deviations were observed in the CBCT STL models (Figure 3),

with mean deviations ranging between 0.63 mm and 0.80 mm. Liang et al. (2010) reported

mean deviations of between 0.165 and 0.386 mm for CBCT STL models [17], though

deviations of up to 2-5 mm have also been reported [18]. The overall low accuracy of

the CBCT STL models found in this study could be caused by the relatively high amount of

noise, artefacts [19], and inhomogeneities in the CBCT X-ray beam [20].

The geometrical deviations that were observed in the orbital area of all STL models

could have been caused by the thin (less than 1 mm) and unique anatomy of the orbits.

Currently, most CT scanners have an in-plane spatial resolution of approximately 0.3

mm and a slice thickness of 0.6 mm and can therefore not detect such bony structures

(smaller than 0.6 mm). Moreover, the complex, scaffold-like structure of the orbital area

also comprises thick supportive bony structures. These structures attenuate the lower part

INFLU

EN

CE

OF C

T PAR

AM

ETE

RS O

N STL M

OD

EL A

CC

UR

AC

Y

38

3

Figure 6. Beam hardening and photon starvation in the orbital area can cause an enlargement of the STL model when compared to the actual bone.

of the radiation energy spectrum and lead to beam hardening and photon starvation [21]

(Figure 6). Hardened energy photons are less prone to attenuate in thin bony structures,

causing a loss of contrast in these structures [22]. Interestingly, when compared with

routine MDCT scanning protocols, min kV/min mA and hence low-dose MDCT protocols

resulted in similar geometric deviations and an increase of detail in the orbital area (Figure

5). These results are in good agreement with a previous study by Oka et al. (2009), who

found minor geometrical differences between bone models constructed using normal and

low-dose CT parameters.

Clinical implicationsWhile patient-specific medical constructs can be fabricated with great accuracy (< 0.1 mm)

using AM technologies, their design is more challenging since it is dictated by the correctness

of the STL model. A recent study by Stoor et al. (2014) reported that 24 per cent of

their additively manufactured orbital implants did not fit due to a poor representation of

the orbital walls in the corresponding CT-based STL models. The consequence of ill-fitting

implants is that increased movements can jeopardize wound healing and tissue repair

and subsequently lead to complications after surgery [24]. Furthermore, geometrically

inaccurate implants can cause nerve impingements in complex anatomical areas such as

the orbit, with devastating outcomes. This opens new questions concerning the liability

of AM medical constructs. Therefore, technicians working in the medical field need to

understand the inaccuracies that can be introduced during the CT imaging and STL

conversion process required for the fabrication of medical AM constructs.

Limitations of this studyThe major limitation of this study was that skulls without soft tissues were used.

The photon energy spectrum of an X-ray beam changes as it passes through water [25],

which results in differences in absorption and scattering during clinical CT scanning.

Furthermore, the intensity values of soft tissues are closer to the values of bone, hence

the presence of soft tissues will make bone segmentation more difficult. Therefore,

INFLU

EN

CE

OF C

T PAR

AM

ETE

RS O

N STL M

OD

EL A

CC

UR

AC

Y

39

3

the reported accuracies found in this study are not simply generalizable to clinical

conditions. Nevertheless, dry skulls combined with optical scanning technology offered

an highly accurate and detailed 3D geometric “gold standard” to evaluate the accuracy

of CT-based STL models. The acquisition of a gold standard is not straightforward for

complex anatomical structures, and even impossible to acquire in an vivo study.

CONCLUSIONCT imaging technologies and their acquisition parameters have an effect on the accuracy

of medical AM constructs. MDCT, DECT and CBCT scanners differ in image quality and

STL model accuracy in the maxillofacial area. Overall, MDCT scanners offered the best

image quality and STL models. The use of low-kV and low-mA instead of routine MDCT

protocols can markedly reduce the effective dose in patients who require patient-specific

AM constructs.

INFLU

EN

CE

OF C

T PAR

AM

ETE

RS O

N STL M

OD

EL A

CC

UR

AC

Y

40

3

REFERENCES1. Ventola C L (2014) Medical Applications

for 3D Printing : Current and Projected Uses. Pharm. Ther. 39:704–711

2. Santler G, Karcher H, and Ruda C (1998) Indications and limitations of three-dimensional models in cranio-maxillofacial surgery. J. Cranio-Maxillofacial Surg. 26:11–16

3. D’Urso P S, Barker T M, Earwaker W J, Bruce L J, Atkinson R L, Lanigan M W, Arvier J F, and Effeney D J (1999) Stereolithographic biomodelling in cranio-maxillofacial surgery: a prospective trial. J. Cranio-Maxillofacial Surg. 27:30–37

4. Giannatsis J and Dedoussis V (2009) Additive fabrication technologies applied to medicine and health care: a review. Int. J. Adv. Manuf. Technol. 40:116–127

5. Guarino J, Tennyson S, McCain G, Bond L, Shea K, and King H (2007) Rapid Prototyping Technology for Surgeries of the Pediatric Spine and Pelvis Benefits Analysis. J. Pediatr. Orthop. 27:955–960

6. Parthasarathy J (2014) 3D modeling, custom implants and its future perspectives in craniofacial surgery. Ann. Maxillofac. Surg. 4:9–18

7. Suetens P (2009) Fundamentals of Medical Imaging. Cambridge University Press, New York, NY

8. Tawfik A M, Kerl J M, Razek A A, Bauer R W, Nour-Eldin N E, Vogl T J, and Mack M G (2011) Image quality and radiation dose of dual-energy CT of the head and neck compared with a standard 120-kVp acquisition. Am. J. Neuroradiol. 32:1994–1999

9. Lofthag-Hansen S (2010) Cone Beam Computed Tomography: Radiation dose and image quality assessments, University of Gothenburg, Gothenburg

10. Ludlow J B and Ivanovic M (2008) Comparative dosimetry of dental CBCT devices and 64-slice CT for oral and

maxillofacial radiology. Oral Surgery, Oral Med. Oral Pathol. Oral Radiol. Endodontology 106:106–114

11. Pauwels R, Beinsberger J, Collaert B, Theodorakou C, Rogers J, Walker A, Cockmartin L, Bosmans H, Jacobs R, Bogaerts R, and Horner K (2012) Effective dose range for dental cone beam computed tomography scanners. Eur. J. Radiol. 81:267–271

12. Dawood A, Patel S, and Brown J (2009) Cone beam CT in dental practice. Br. Dent. J. 207:23–28

13. Mallepree T and Bergers D (2009) Accuracy of medical RP models. Rapid Prototyp. J. 15:325–332

14. Thorhauer E, Miyawaki M, Illingworth K, Holmes J A, and Anderst W (2010) Accuracy of bone and cartilage models obtained from CT and MRI. Poster presented at the 34th Annual Meeting of the American Society of Biomechanics, 18–21

15. Rathnayaka K, Momot K I, Noser H, Volp A, Schuetz M A, Sahama T, and Schmutz B (2014) Quantification of the accuracy of MRI generated 3D models of long bones compared to CT generated 3D models. Med. Eng. Phys. 34:357–363

16. Oka K, Murase T, Moritomo H, Goto A, Sugamoto K, and Yoshikawa H (2009) Accuracy analysis of three-dimensional bone surface models of the forearm constructed from multidetector computed tomography data. Int. J. Med. Robot. Comput. Assist. Surg. 5:452–457

17. Liang X, Lambrichts I, Sun Y, Denis K, Hassan B, Li L, Pauwels R, and Jacobs R (2010) A comparative evaluation of Cone Beam Computed Tomography (CBCT) and Multi-Slice CT (MSCT). Part II: On 3D model accuracy. Eur. J. Radiol. 75:270–274

18. Shahbazian M, Jacobs R, Wyatt J, Willems G, Pattijn V, Dhoore E, Van

INFLU

EN

CE

OF C

T PAR

AM

ETE

RS O

N STL M

OD

EL A

CC

UR

AC

Y

41

3

Lierde C, and Vinckier F (2010) Accuracy and surgical feasibility of a CBCT-based stereolithographic surgical guide aiding autotransplantation of teeth: In vitro validation. J. Oral Rehabil. 37:854–859

19. Pauwels R, Nackaerts O, Bellaiche N, Stamatakis H, Tsiklakis K, Walker A, Bosmans H, Bogaerts R, Jacobs R, and Horner K (2013) Variability of dental cone beam CT grey values for density estimations. Br. J. Radiol. 86:20120135

20. Feldkamp L A, Davis L C, and Kress J W (1984) Practical cone-beam algorithm. J. Opt. Soc. Am. 1:612

21. Zbijewski W, Stayman J, and Muhit A (2012) CT Reconstruction Using Spectral and Morphological Prior Knowledge: Application to Imaging the Prosthetic Knee. The second international conference on image formation in X-ray computed tomography, Salt Lake City, UT, 434–438

22. Koivisto J, Wolff J, Järnstedt J, Dastidar P, and Kortesniemi M (2014) Assessment of the effective dose in supine, prone, and oblique positions in the maxillofacial region using a novel combined extremity and maxillofacial cone beam computed tomography scanner. Oral Surg. Oral Med. Oral Pathol. Oral Radiol. 118:355–362

23. Stoor P, Suomalainen A, Lindqvist C, Mesimäki K, Danielsson D, Westermark A, and Kontio R K (2014) Rapid prototyped patient specific guiding implants for orbital wall defects. J. Cranio-Maxillofacial Surg. 42:1644–1649

24. Lienau J, Schell H, Duda G N, Seebeck P, Muchow S, and Bail H J (2005) Initial vascularization and tissue differentiation are influenced by fixation stability. J. Orthop. Res. 23:639–645

25. Barrett J F and Keat N (2004) Artifacts in CT: Recognition and Avoidance. RadioGraphics ` 24:1679–1691

4 IMPACT OF PRONE, SUPINE AND OBLIQUE PATIENT POSITIONING

ON CBCT IMAGE QUALITY, CONTRAST TO NOISE RATIO

AND FIGURE OF MERIT VALUE IN THE MAXILLOFACIAL REGION

Juha Koivisto, Maureen van Eijnatten, Jorma Järnstedt,

Kisri Holli-Helenius, Prasun Dastidar, Jan Wolff

Dentomaxillofacial Radiology (2017). 46 (6): 20160418doi: 10.1259/dmfr.20160418

IMPA

CT O

F PRO

NE

, SUPIN

E A

ND

OB

LIQU

E PA

TIEN

T POSITIO

NIN

G O

N C

BC

T IMA

GE

QU

ALITY

44

4

ABSTRACTObjectiveTo assess the impact of supine, prone and oblique patient imaging positions on the image

quality, contrast-to-noise ratio and figure of merit value in the maxillofacial region using

a cone beam computed tomography scanner. Furthermore, the CBCT supine images were

compared to supine MSCT images.

MethodsOne fresh frozen cadaver head was scanned in prone, supine and oblique imaging

positions using a mobile CBCT scanner. MSCT images of the head were acquired in

a supine position. Two radiologists graded the CBCT and MSCT images at ten different

anatomical sites according to their image quality using a six-point scale. CNR and FOM

values were calculated at two different anatomical sites on the CBCT and MSCT images.

ResultsThe best image quality was achieved in prone imaging position for sinus, mandible

and maxilla, followed by supine and oblique imaging positions. 12-mA prone images

presented high delineation scores for all anatomical landmarks, except for the ear region

(carotid canal), which presented adequate to poor delineation scores for all studied head

positions and exposure parameters. The MSCT scanner offered similar image qualities to

the 7.5-mA supine images acquired using the mobile CBCT scanner. The prone imaging

position offered the best CNR and FOM values on the mobile CBCT scanner.

ConclusionsHead positioning has an impact on CBCT image quality. The best CBCT image quality can

be achieved using the prone and supine imaging positions. The oblique imaging position

offers inadequate image quality except in the sinus region.

IMPA

CT O

F PRO

NE

, SUPIN

E A

ND

OB

LIQU

E PA

TIEN

T POSITIO

NIN

G O

N C

BC

T IMA

GE

QU

ALITY

45

4

INTRODUCTION Mobile cone beam computed tomography (CBCT) scanners were first introduced in the field

of maxillofacial surgery in 2005 [1]. Since then, mobile CBCT scanners [2]–[8] have become

increasingly popular as they offer comparable images to those acquired on traditional

multislice computed tomography (MSCT) scanners [9]. The major advantages of CBCT

scanners are their lower cost, size, weight and effective dose [10]–[12]. The lower effective

dose is the result of the default exposure parameters and the partial gantry rotation used

in CBCT scanners when compared to MSCT scanners [13], [14]. It has, however, been

reported that the partial CBCT gantry rotation produces a non-uniform dose distribution

[15] that can in some cases affect image quality. Nevertheless, the partial CBCT gantry

rotation offers acceptable diagnostic image quality at a lower effective dose than MSCT

scanners [16].

A recent study by Koivisto et al. [17] reported the positive effect of head positioning

on the effective dose in CBCT scanners. Furthermore, a study by Yan et al. [18] reported

a close relationship between image quality and effective dose when using low dose CBCT

protocols. Taking the results of the previous studies into account, it can be hypothesized

that CBCT images could be optimised by choosing the correct patient imaging position.

Currently, MSCT and CBCT image quality assessments are based on the visual

grading of predefined anatomical landmarks [19]–[21], the quantitative analysis of

physical properties such as modulation transfer function (MTF) [22] and contrast-to-noise

ratio (CNR) [23]. The image contrast is affected by the x-ray mean photon energy that

subsequently has an effect on the image quality [24]. Moreover, the relationship between

the image quality and patient dose is commonly quantified using the traditional figure of

merit (FOM) value that is defined as the (CNR) [2] divided by the incident patient exposure

[24]–[27]. The figure of merit value can provide useful information on the benefits and

drawbacks of different imaging positions in CBCT and MSCT scanners. It also provides

a means of comparing the ratio between the image quality and the resulting effective

dose. Furthermore the FOM offers the possibility of calculating a numerical value that

can be used to compare two inherently different imaging modalities. Moreover, the FOM

value can be used to select an appropriate protocol depending on the diagnostic task.

The combination of CNR and FOM calculations with image quality assessments offer an

adequate method for comparing and grading MSCT and CBCT images.

The objectives of this study were to assess the impact of three different patient imaging

positions (supine, prone, oblique) on CBCT image quality, contrast-to-noise ratio and

the figure of merit value in the maxillofacial region. The second objective was to compare

the CBCT supine images with MSCT images acquired in the same position.

MATERIALS AND METHODSThis study was performed according to the Ethical Principles for Medical Research Involving

Human Subjects as defined by the World Medical Association (Helsinki Declaration) [28].

IMPA

CT O

F PRO

NE

, SUPIN

E A

ND

OB

LIQU

E PA

TIEN

T POSITIO

NIN

G O

N C

BC

T IMA

GE

QU

ALITY

46

4

One fresh frozen human cadaver head was provided by the Department of Anatomy,

Tampere University Hospital for image quality assessment.

ScannersCadaver images were obtained using a Planmed Verity mobile CBCT scanner (Planmed

Oy, Helsinki, Finland) and a Philips Brilliance-64 (64 slice) MSCT scanner (Philips Medical

Systems, Eindhoven, Netherlands). The mobile CBCT scanner uses a partial 210º gantry

scan and 3.0 mm Al and 0.5 mm Cu filters. The Philips Brilliance 64 MSCT scanner

uses a 360º rotation scan and combined filtration consisting of 2.5 mm aluminum and

1.2 mm titanium.

The mobile scanner was initially developed for extremity imaging and has recently been

approved by the American Food and Drug Administration (FDA) for maxillofacial imaging.

The scanner is equipped with a head positioning tray and a detachable forehead support.

During oblique imaging, the cadaver head was placed on the default tray and attached to

the forehead support. In order to acquire prone and supine images, the forehead support

was removed and the gantry was tilted accordingly.

Imaging positions A total of nine CBCT images were acquired in supine (Figure 1B), prone (Figure 1C) and

oblique (Figure 1D) imaging positions. Furthermore, one MSCT image was acquired using

the default supine imaging position (Figure 1A).

Exposure protocolsCBCT images were obtained using a constant 96 kV tube voltage. Three different tube

currents (1): 4.8 mA, (2): 7.5 mA and (3): 12 mA were used to encompass the full dynamic

Figure 1. A schematic presentation of the MSCT (A) and mobile CBCT (B, C, D) scanner used in this study with different head positions: supine (A, B), prone (C), and oblique (D).

IMPA

CT O

F PRO

NE

, SUPIN

E A

ND

OB

LIQU

E PA

TIEN

T POSITIO

NIN

G O

N C

BC

T IMA

GE

QU

ALITY

47

4

range of the mobile CBCT scanner. All CBCT exposures were performed using the smallest

possible voxel size (0.2 mm). The MSCT exposures were performed using a default standard

high-resolution imaging protocol (120 kV, 93 mA).

All effective doses required to calculate the figure of merit (FOM) value were scaled

proportionally using the exposure values (mAs) attained in a previous study by Koivisto

et al. [17] using an anthropomorphic phantom. All exposure protocols, parameters and

calculated effective doses are presented in Table 1.

Standard axial, coronal and sagittal reconstructions were performed on a Philips

Brilliance Extended Workstation V 4.5.2.40007 (Philips Healthcare, the Netherlands).

Table 1. Exposure parameters of the CBCT and MSCT exposures

Planmed Verity Philips Brilliance

CBCT 64 MSCT

Imaging position Supine/ Supine/ Supine/   SupineProne/ Prone/ Prone/ -

  Oblique Oblique Oblique -

Imaging protocol Protocol Protocol Protocol Protocol  4.8 mA 7.5 mA 12 mA High Resolution

Tube voltage (kVp) 96 96 96 120Tube current (mA) 4.8 7.5 12 93

Exposure time (s) 6 6 6 0.912

Q (mAs) 28.8 45 72 85

Slice thickness (mm) 0.2 0.2 0.2 0.67

Slice increment (mm) 0.2 0.2 0.2 0.33

Pitch (mm) NA NA NA 0.55

CTDIvol (mGy) - - - 11

Voxel size (mm) 0.2 x 0.2 x 0.2 0.2 x 0.2 x 0.2 0.2 x 0.2 x 0.2 0.33x0.352x0.352

Scan angle 210º 210º 210º 360º

Frame number 300 300 300 -

Radiation source Pulsed Pulsed Pulsed Continuous

Field of view (cm) 13 x 16 13 x 16 13 x 16 16 x 18

Eff.dose, Supine (µSv)* 125 195 312 781Eff.dose, Prone (µSv)* 97 151 242 -

Eff.dose, Oblique (µSv)* 29 45 72 -

NA, not applicable; CTDI, computed tomography dose index. *Calculated from the effective doses reported by Koivisto et al. 2014

IMPA

CT O

F PRO

NE

, SUPIN

E A

ND

OB

LIQU

E PA

TIEN

T POSITIO

NIN

G O

N C

BC

T IMA

GE

QU

ALITY

48

4

The reconstructed image data consisting of 474 MSCT and 652 CBCT DICOM (Digital

Imaging and Communications in Medicine) files were subsequently exported to an Agfa

Impax PACS system (Agfa Healthcare, Mortsel, Belgium) and randomly visually graded

by two experienced radiologists using the same viewing conditions. All evaluations were

repeated after four weeks to assess the intra-observer reliability.

Image quality assessment All image quality assessments were performed on ten different anatomical landmarks in

four different anatomical regions: sinus (blue), ear (yellow), maxilla (green) and mandible

(red) (Figure 2) The image quality was graded by two radiologists using the following

scale: Level 5; excellent delineation of structures and excellent image quality, Level 4;

clear delineation of structures and excellent image quality, Level 3; anatomic structures

still fully assessable in all parts and good image quality, Level 2: structures identified and

results in adequate image quality, Level 1; anatomic structures not identifiable due to poor

image quality, Level 0; no diagnostic value.

Statistical analysis of the image quality The inter-rater agreement was evaluated by calculating the percentage of absolute

agreement and two-way random effects intra-class correlation coefficients (ICC) with

absolute agreement corresponding to 95% confidence interval (CI) [29]. The interclass

Figure 2. Anatomical landmarks assessed on the reconstructed MSCT and CBCT images

IMPA

CT O

F PRO

NE

, SUPIN

E A

ND

OB

LIQU

E PA

TIEN

T POSITIO

NIN

G O

N C

BC

T IMA

GE

QU

ALITY

49

4

correlation coefficient (ICC) (Two-Way Mixed) model (2,2) was used to determine inter-

rater reliability [30].

The intra-rater reliability was calculated for four selected protocols (oblique position

(4.8 mA), supine position (12 mA) and prone position (7.5 mA)) using ICC (Two-Way Mixed)

model (3,1) [30]. ICCs range from 0 with no agreement to +1 with perfect agreement in

the interpretation of the ICC. Landis et al. [31] has characterized the values of reliability

coefficients as follows: less than 0.4, poor; 0.4 to 0.75, fair to good; and greater than 0.75,

excellent. Although arbitrary, these divisions provide useful benchmarks [29].

Contrast-to-noise ratioThe contrast-to-noise ratio (CNR) evaluation was performed in the fossa olfactorius

(anatomical landmark no. 6) and the apical aspect of tooth 43 (D43, anatomical landmark

no. 2). These anatomical sites were chosen since they represent commonly diagnosed

regions in the head area. The locations of ROIs were verified by superimposing CBCT and

MSCT images. During superimposition, all CBCT and corresponding MSCT images were

manually aligned into a coordinate system using control points chosen by the experienced

radiologist [32]–[34].

The CNR was calculated using the following equation:

(1)

where, SROI is the mean signal at the anatomical landmark in grey value unit, SBGR is

the mean signal in the background (HU) and σmean is the average standard deviation in

the anatomical landmark and background ROIs (Figure 3). Furthermore, the correlation

between CNR and the subjective image quality was calculated to assess the association

between the two image quality indicators.

The intra-rater reliability was calculated for four selected protocols (oblique position (4.8 mA), supine

position (12 mA) and prone position (7.5 mA)) using ICC (Two-Way Mixed) model (3,1) [30]. ICCs

range from 0 with no agreement to +1 with perfect agreement in the interpretation of the ICC. Landis et

al. [31] has characterized the values of reliability coefficients as follows: less than 0.4, poor; 0.4 to 0.75,

fair to good; and greater than 0.75, excellent. Although arbitrary, these divisions provide useful

benchmarks [29].

Contrast-to-noise ratio

The contrast-to-noise ratio (CNR) evaluation was performed in the fossa olfactorius (anatomical

landmark no. 6) and the apical aspect of tooth 43 (D43, anatomical landmark no. 2). These anatomical

sites were chosen since they represent commonly diagnosed regions in the head area. The locations of

ROIs were verified by superimposing CBCT and MSCT images. During superimposition, all CBCT and

corresponding MSCT images were manually aligned into a coordinate system using control points

chosen by the experienced radiologist [32]–[34].

The CNR was calculated using the following equation:

(1)

where, SROI is the mean signal at the anatomical landmark in grey value unit, SBGR is the mean signal in

the background (HU) and σmean is the average standard deviation in the anatomical landmark and

background ROIs (Figure 3). Furthermore, the correlation between CNR and the subjective image

quality was calculated to assess the association between the two image quality indicators.

Figure 3. Axial view of the tooth 43 (apex, D43) and background locations used for CNR calculation for MSCT

supine position (A) and for CBCT using supine (B), prone (C) and oblique (D) positions.

CNR SROI SBGR

mean

Figure 3. Axial view of the tooth 43 (apex, D43) and background locations used for CNR calculation for MSCT supine position (A) and for CBCT using supine (B), prone (C) and oblique (D) positions.

IMPA

CT O

F PRO

NE

, SUPIN

E A

ND

OB

LIQU

E PA

TIEN

T POSITIO

NIN

G O

N C

BC

T IMA

GE

QU

ALITY

50

4

Figure of MeritIn this study, the figure of merit (FOM) value was used to assess the diagnostic efficacy of

the image quality versus the effective dose obtained using different imaging protocols.

The FOM value was calculated using the following equation described by Ogden et al. [26]

(2)

where E is the effective dose. In the present study, the FOM value was calculated for two

anatomical landmarks: the fossa olfactorius and the apex of tooth 43.

RESULTSImage quality assessmentThe average image quality results of nine different CBCT imaging protocols acquired in

different imaging positions and one MSCT protocol acquired in one position are presented

in Figure 4.

Statistical analysis of image qualityStatistical analysis was performed on all CBCT and MSCT image quality results.

The absolute agreement percentages, inter-rater intraclass correlation coefficients (ICC)

and 95% confidence intervals (CI) of all examination protocols are presented in Table 2.

The intra-rater reliability results of the two observers are presented in Table 3.

Figure of Merit

In this study, the figure of merit (FOM) value was used to assess the diagnostic efficacy of the image

quality versus the effective dose obtained using different imaging protocols. The FOM value was

calculated using the following equation described by Ogden et al. [26]

(2)

where E is the effective dose. In the present study, the FOM value was calculated for two anatomical

landmarks: the fossa olfactorius and the apex of tooth 43.

Results

Image quality assessment

The average image quality results of nine different CBCT imaging protocols acquired in different

imaging positions and one MSCT protocol acquired in one position are presented in Figure 4.

Figure 4. Results of subjective image quality assessments undertaken by two radiologists using different CBCT

and MSCT imaging protocols.

FOM CNR2

E

Figure 4. Results of subjective image quality assessments undertaken by two radiologists using different CBCT and MSCT imaging protocols.

IMPA

CT O

F PRO

NE

, SUPIN

E A

ND

OB

LIQU

E PA

TIEN

T POSITIO

NIN

G O

N C

BC

T IMA

GE

QU

ALITY

51

4

Contrast-to-noise ratioThe mean CBCT contrast-to-noise values calculated in different imaging positions were as

follows: 3.6 in the prone imaging position, 1.8 in the supine imaging position and 1.5 in

the oblique imaging position. The Philips Brilliance 64 MSCT scanner mean CNR was 2.6.

Furthermore, the correlation between CNR and subjective image quality were 0.8217 for

tooth 43 and 0.3581 for the fossa olfactorius. The CNR values and the subjective image

qualities of the apex of tooth 43 and the fossa olfactorius are presented in Figure 5A and

Figure 5B, respectively.

Figure of MeritThe mean FOM values of the Planmed Verity CBCT scanner in different imaging positions

were as follows: 96.1 in the prone imaging position, 24.3 in the supine imaging position

and 22.3 in the oblique imaging position. The mean FOM value of the Philips Brilliance 64

MSCT scanner was 10.6. The FOM values of the apex of tooth 43 and the fossa olfactorius

obtained using one MSCT modality and multiple CBCT imaging modalities are presented

in Figure 6A and Figure 6B, respectively.

Table 2. Inter-rater absolute agreement (%), intra-class correlation (ICC) and confidence interval at 95% confidence level

Protocol Absolute agreement (%) ICC 95% Confidence Interval (CI)

MSCT (supine) 90 0.942 0.836, 0.984Oblique (12 mA) 90 0.967 0.908, 0.991Oblique (7.5 mA) 70 0.955 0.868, 0.988Oblique (4.8 mA) 90 0.964 0.897, 0.991Supine (12 mA) 80 0.893 0.696, 0.971Supine (7.5 mA) 70 0.880 0.659, 0.967Supine (4.8 mA) 70 0.887 0.675, 0.969Prone (12 mA) 70 0.889 0.687, 0.970Prone (7.5 mA) 70 0.898 0.710, 0.972Prone (4.8 mA) 80 0.940 0.827, 0.984

Table 3. Intra-rater reliability of selected protocols 

 Philips Brilliance 64 MSCTSupine 93 mA  

Planmed Verity CBCT

Oblique 4.8 mA Supine 12 mA Prone 7.5 mA

Observer 1 0.780 0.873 0.571 0.856

Observer 2 1.000   0.877 0.667 0.816

IMPA

CT O

F PRO

NE

, SUPIN

E A

ND

OB

LIQU

E PA

TIEN

T POSITIO

NIN

G O

N C

BC

T IMA

GE

QU

ALITY

52

4

DISCUSSIONIn this study, the impact of supine, prone and oblique imaging positioning on image quality

in four different skull regions was assessed. Furthermore, the contrast-to-noise ratio (CNR)

and the figure of merit value were calculated and compared. Finally, the image quality of

CBCT images acquired in the supine imaging position were compared to MSCT images

acquired in the same imaging position.

The combined mobile extremity/maxillofacial CBCT scanner used in this study is unique in

that it uses an adjustable gantry instead of a fixed gantry [35]. The adjustable gantry which

moves in an enclosed housing, offers the possibility of acquiring patient images in seated

and standing positions. Furthermore, supine, prone and oblique imaging positions are

possible. The position of the head with respect to the X-ray beam and reciprocal sensor in

the mobile CBCT scanner is almost identical to other dental CBCT scanners on the market

in the supine (Figure 2B) and prone (Figure 2C) imaging positions. Therefore, there are no

Figure 5. The image quality and CNR values of tooth 43 (A) and the fossa olfactorius (B) using different imaging positions and protocols.

Figure 6. The figure of merit values of tooth 43 (A) and fossa olfactorius (B) using different imaging positions and protocols.

IMPA

CT O

F PRO

NE

, SUPIN

E A

ND

OB

LIQU

E PA

TIEN

T POSITIO

NIN

G O

N C

BC

T IMA

GE

QU

ALITY

53

4

major image quality differences between mobile CBCT scanners and other dental CBCT

scanners. However, the results found in this study concerning the oblique imaging position

are unique to the mobile CBCT scanner and therefore do not apply to other dental

CBCT scanners.

CBCT image qualitySinus: In the sinus region, the differences between the image qualities acquired using

the supine and oblique imaging positions were negligible. However, the best CBCT mean

image quality was observed in the prone imaging position. When compared to the prone

and oblique imaging positions, the supine imaging position offered only poor image

quality in the fossa olfactorius using 4.8 mA tube current. Furthermore, inadequate image

quality was observed in the fossa olfactorius using the prone and supine imaging positions

and 7.5 mA tube current. In the sinus region, only minor variations in the image quality

were observed when using 4.8 mA and 7.5 mA. However, when using the highest 12 mA

tube current, a significantly better image quality was observed in the sinus area. This

phenomenon could be caused by the better signal-to-noise ratio attained using 12 mA

tube current. Moreover, the better image quality observed in the prone imaging position

could be due to the closer proximity of the cadaver head to the scanner sensor.

The low image quality achieved in the supine imaging position can be explained

by the stronger attenuation and scattering resulting from the thick and complex bony

structures in the anterior part of the skull [32], [33]. Furthermore, the low image quality

observed in the oblique imaging position is likely to be caused by the longer radiation path

length and subsequent attenuation caused by the parietal bones at the top of the cranium

[22], [34]. These findings suggest that it could be advantageous to use the oblique imaging

position instead of the supine imaging position in the sinus region. This hypothesis is in

good agreement with a previous study by Heiland et al. who reported on the advantage

of using the oblique imaging position to image the sinuses on a Siremobil Iso-C3D CBCT

scanner [1]. However it must be noted that an oblique imaging position is not attainable

using a standard upright CBCT scanner.

Ear: In the ear region, the oblique imaging position resulted in poor image quality

of the carotid canal. Furthermore, the supine and prone imaging positions offered only

adequate image quality of the carotid canal. The low image quality observed in the oblique

position could have been caused by the attenuation of the parietal bones situated on

the top of the cranium.

The difficulty of acquiring adequate image quality in the carotid canal regardless of

the tube current (4.8 mA to 12 mA) could be due to the attenuation caused by the dense

bony structures of the temporal bone. These findings are in good agreement with a previous

study by Pereira et al. who reported on the difficulty of obtaining adequate images in

the TMJ region due to the dense bony structures in the petrous part of the temporal bone

[36]. These findings are further supported by the results of Kwong et al. [37] and Sohaib

IMPA

CT O

F PRO

NE

, SUPIN

E A

ND

OB

LIQU

E PA

TIEN

T POSITIO

NIN

G O

N C

BC

T IMA

GE

QU

ALITY

54

4

et al. [38] who concluded that no significant image quality difference occurred when tube

currents were changed.

Maxilla: Using the mobile CBCT scanner, the prone and supine imaging positions

offered better image quality in the maxilla than the oblique imaging position. The prone

imaging position offered the best image quality of tooth 16 and the foramen infraorbitale

using 12 mA. The impact of different tube currents on image quality in all imaging positions

was moderate.

Mandible: In the mandible, the prone and supine imaging positions resulted in good

image quality for all three investigated tube currents. Furthermore, the oblique imaging

position demonstrated an adequate image quality of the apex of tooth 43 and poor image

quality of the foramen mentale using 7.5 mA tube current. However, only poor image

qualities were attainable in the mandible using the oblique imaging position and 4.8 mA

tube current. The impact of different tube currents on the image quality in the prone

imaging position was negligible and moderate in the oblique imaging position. However,

the tube current did not have any effect on image quality in the supine imaging position.

In general, the low image quality observed in this study using the oblique imaging

position is likely due to the anatomy of the head. Moreover, in the oblique position, the x-ray

beam enters the head on the back side of the cranium, and therefore the radiation path

length through the head becomes longer than in the prone position. This phenomenon

subsequently causes stronger attenuation, which has a negative effect on the overall

image quality [17].

Comparison between MSCT and CBCT images acquired in the supine positionThe MSCT, CBCT inter-rater and intra-rater reliabilities were assessed according to

the characterization method decribed by Landis et al. [31] The MSCT intra-rater reliability

was excellent. On the other hand, the CBCT intra-rater reliability was excellent in

the oblique and prone imaging positions and only good in the supine position.

In the supine position, the mobile CBCT scanner offered similar image quality to that

of the MSCT scanner when using 7.5 mA tube current. However, the mobile CBCT scanner

resulted in lower image quality and higher local variations in image quality when a 4.8 mA

tube current was used. Furthermore, when using 12 mA tube current, the mobile CBCT

scanner offered a good delineation of most anatomical structures and better image quality

than those acquired on the MSCT scanner.

Contrast-to-noise ratioIn this study, the best CNR value was observed in the apex of tooth 43 in the prone

imaging position. The lower CNR values observed in the oblique imaging position could

have been caused by the longer radiation path length and the parietal bones on the top

of the cranium. Interestingly, in the oblique imaging position, the delineation of the fossa

IMPA

CT O

F PRO

NE

, SUPIN

E A

ND

OB

LIQU

E PA

TIEN

T POSITIO

NIN

G O

N C

BC

T IMA

GE

QU

ALITY

55

4

olfactorius resulted in a higher CNR value than in the supine imaging position. The lower

CNR value observed in the fossa olfactorius using the supine imaging position can be

explained by the stronger attenuation and scattering from the thick and complex bony

structures in the anterior nasofrontal area of the skull. [39] Furthermore, the correlation

between the CNR value and subjective image quality results was excellent in the apex of

tooth 43. However, only a fair correlation between the CNR value and image quality was

achieved in the fossa olfactorius. These results are in good agreement with a recent study

by Choi et al. [22] who described a strong association between subjective image quality

and CNR values.

Figure of MeritThe prone imaging position offered the best figure of merit value in the fossa olfactorius

and the apex of tooth 43. The differences between the oblique and supine imaging

positions were negligible. However, in the fossa olfactorius, the differences in the FOM

values attained in the oblique and supine imaging positions were significant. This could

have been caused by the highest CNR value attained at a moderate effective dose.

The MSCT scanner resulted in a higher FOM value than the mobile CBCT scanner in

the supine imaging position. However, the figure of merit value obtained using the MSCT

was lower than those acquired using the mobile CBCT scanner in prone and oblique

imaging positions.

The lower figure of merit value obtained on the MSCT scanner can be explained by

the manufacturer-recommended high-resolution imaging parameters that subsequently

resulted in a higher effective dose than in the mobile CBCT scanner. Nevertheless,

Table 4. A comparison between the MSCT and CBCT image qualities acquired in supine imaging positions.

RegionAnatomical landmark

Planmed Verity CBCT

 

Philips Brilliance 64 MSCT

4.8 mA 7.5 mA 12 mA High Resolution

Sinus Fossa olfact. poor poor good AdequateLamina.papyr. adequate adequate adequate AdequateHiatus max. adequate adequate good AdequateMiddle turb. adequate adequate good Adequate

  Ostium adequate adequate good   Adequate

Ear Carotid canal poor adequate adequate   Adequate

Maxilla Foramen inf. adequate good good   Good  D 16 apex adequate good good   Adequate

Mandible D 43 apex good good good   Adequate  Foramen.ment. good good good   Good

IMPA

CT O

F PRO

NE

, SUPIN

E A

ND

OB

LIQU

E PA

TIEN

T POSITIO

NIN

G O

N C

BC

T IMA

GE

QU

ALITY

56

4

the effective dose can be significantly reduced by using a lower mAs value that subsequently

does not markedly affect the diagnostic image quality in the sinus region [38].

One weakness of this study is the use of a frozen cadaver head for image quality

assessment During freezing residual water in the cadavar expands and subequently causes

anatomical alterations especially in the fine bony structures of the sino-nasal region.

Furthermore, frozen fatty acids in the brain [40] decrease in volume and consequently

increase the attenuation resulting in an increase of the HU- or the grey-values [40], [41].

The aforementioned factors subsequently reduce the image quality and contrast resolution

[42]. Moreover, the figure of merit value was assessed based on the CNR value from

a cadaver head while the effective dose used for calculating the FOM value was obtained

using an anthropomorphic phantom. Another drawback of this study is that the image

quality was assessed for only one anatomical landmark in the ear region while there is

a growing interest for diagnostic imaging of temporal bones due to the cochlear implants

[43]. Furthermore, an interpretation with the results obtained using the MSCT scanner

should be carried out with care since the images were acquired using standard exposure

parameters without image quality optimisation.

CONCLUSIONSHead imaging positioning has an impact on the overall CBCT image quality. The best

mean CBCT image quality, CNR value and FOM value were observed using the prone

imaging position for all diagnostic regions. The oblique imaging position offers inadequate

image quality except in the sinus region. The MSCT scanner offered similar mean image

quality in the supine position to that acquired on the mobile CBCT scanner using 7.5 mA

tube current.

CONFLICTS OF INTEREST Juha Koivisto is an employee of Planmeca Oy.

IMPA

CT O

F PRO

NE

, SUPIN

E A

ND

OB

LIQU

E PA

TIEN

T POSITIO

NIN

G O

N C

BC

T IMA

GE

QU

ALITY

57

4

REFERENCES1. Heiland M, Schulze D, Blake F, and

Schmelzle R (2005) Intraoperative imaging of zygomaticomaxillary complex fractures using a 3D C-arm system. Int. J. Oral Maxillofac. Surg. 34:369–375

2. Koskinen S K, Haapamaki V V, Salo J, Lindfors N C, Kortesniemi M, Seppala L, and Mattila K T (2013) CT arthrography of the wrist using a novel, mobile, dedicated extremity cone-beam CT (CBCT). Skeletal Radiol. 42:649–657

3. Tuominen E K J, Kankare J, Koskinen S K, and Mattila K T (2013) Weight-bearing CT imaging of the lower extremity. Am. J. Roentgenol. 200:146–148

4. Zbijewski W and Stayman J W (2009) Volumetric soft tissue brain imaging on xCAT, a mobile flat-panel x-ray CT system. Proc. SPIE Med. Imaging. 7258:72582K–72582K–5

5. Zhang J, Weir V, Fajardo L, Lin J, Hsiung H, and Ritenour E R (2009) Dosimetric characterization of a cone-beam O-arm imaging system. J. Xray. Sci. Technol. 17:305–317

6. Siewerdsen J H, Daly M J, Bachar G, Moseley D J, Bootsma G, Brock K K, Ansell S, Wilson G A, Chhabra S, Jaffray D A, and Irish J C (2007) Multimode C-arm fluoroscopy, tomosynthesis, and cone-beam CT for image-guided interventions: from proof of principle to patient protocols. Proc. SPIE. 6510: 65101A–65101A–11

7. Siewerdsen J H, Daly M J, Chan H, Nithiananthan S, Hamming N, Brock K K, and Irish J C (2009) High-performance intraoperative cone-beam CT on a mobile C-arm: an integrated system for guidance of head and neck surgery. Proc. SPIE. 7261:72610J–72610J–8

8. Siewerdsen J H (2011) Cone-Beam CT with a Flat-Panel Detector: From Image Science to Image-Guided Surgery. Nucl. Instrum. Methods Phys. Res. A. 648:S241–S250

9. Schafer S, Nithiananthan S, Mirota D J, Uneri A, Stayman J W, Zbijewski W, Schmidgunst C, Kleinszig G, Khanna A J, and Siewerdsena J H (2011) Mobile C-arm cone-beam CT for guidance of spine surgery: image quality, radiation dose, and integration with interventional guidance. Med. Phys. 38:4563–4574

10. Siewerdsen J H, Chan Y, Rafferty M A, Moseley D J, Jaffray D A, and Irish J C (2005) Cone-beam CT with a flat-panel detector on a mobile C-arm: preclinical investigation in image-guided surgery of the head and neck. Proc. SPIE. 5744:789–797

11. Loubele M, Bogaerts R, Van Dijck E, Pauwels R, Vanheusden S, Suetens P, Marchal G, Sanderink G, and Jacobs R (2009) Comparison between effective radiation dose of CBCT and MSCT scanners for dentomaxillofacial applications. Eur. J. Radiol. 71:461–468

12. Ludlow J B and Ivanovic M (2008) Comparative dosimetry of dental CBCT devices and 64-slice CT for oral and maxillofacial radiology. Oral Surgery, Oral Med. Oral Pathol. Oral Radiol. Endodontology. 106:106–114

13. Nemtoi A, Czink C, Haba D, and Gahleitner A (2013) Cone beam CT: a current overview of devices. Dentomaxillofac. Radiol. 42:20120443

14. Kim S, Yoo S, Yin F-F, Samei E, and Yoshizumi T (2010) Kilovoltage cone-beam CT: comparative dose and image quality evaluations in partial and full-angle scan protocols. Med. Phys. 37:3648–3659

15. Fahrig R, Dixon R, Payne T, Morin R L, Ganguly A, and Strobel N (2006) Dose and image quality for a cone-beam C-arm CT system. Med. Phys. 33:4541–4550

16. Lofthag-Hansen S, Thilander-Klang A, and Grondahl K (2011) Evaluation of subjective image quality in relation to diagnostic task for cone beam computed

IMPA

CT O

F PRO

NE

, SUPIN

E A

ND

OB

LIQU

E PA

TIEN

T POSITIO

NIN

G O

N C

BC

T IMA

GE

QU

ALITY

58

4

tomography with different fields of view. Eur. J. Radiol. 80:483–488

17. Koivisto J, Wolff J, Järnstedt J, Dastidar P, and Kortesniemi M (2014) Assessment of the effective dose in supine, prone, and oblique positions in the maxillofacial region using a novel combined extremity and maxillofacial cone beam computed tomography scanner. Oral Surg. Oral Med. Oral Pathol. Oral Radiol. 118:355–362

18. Yan H, Cervino L, Jia X, and Jiang S B (2012) A comprehensive study on the relationship between the image quality and imaging dose in low-dose cone beam CT. Phys. Med. Biol. 57:2063–2080

19. Heiland M, Schulze D, Adam G, and Schmelzle R (2003) 3D-imaging of the facial skeleton with an isocentric mobile C-arm system (Siremobil Iso-C3D). Dentomaxillofac. Radiol. 32:21–25

20. Schulze D, Heiland M, Blake F, Rother U, and Schmelzle R (2005) Evaluation of quality of reformatted images from two cone-beam computed tomographic systems. J. cranio-maxillo-facial Surg. 33:19–23

21. Kamburoglu K, Murat S, Kolsuz E, Kurt H, Yuksel S, and Paksoy C (2011) Comparative assessment of subjective image quality of cross-sectional cone-beam computed tomography scans. J. Oral Sci. 53:501–508

22. Choi J-W, Lee S-S, Choi S-C, Heo M-S, Huh K-H, Yi W-J, Kang S-R, Han D-H, and Kim E-K (2015) Relationship between physical factors and subjective image quality of cone-beam computed tomography images according to diagnostic task. Oral Surg. Oral Med. Oral Pathol. Oral Radiol. 119:357–365

23. Yu L, Vrieze T J, Bruesewitz M R, Kofler J M, DeLone D R, Pallanch J F, Lindell E P, and McCollough C H (2010) Dose and image quality evaluation of a dedicated cone-beam CT system for high-contrast neurologic applications. AJR. Am. J. Roentgenol. 194:W193-201

24. Martin C J (2007) The importance of radiation quality for optimisation in radiology. Biomedical Imaging and Intervention Journal, 3:e38

25. Zamenhof R G (1982) The optimization of signal detectability in digital fluoroscopy. Med. Phys. 9:688–694

26. Ogden K M, Huda W, Khorasani M R, and Scalzetti E M (2008) A comparison of three CT voltage optimization strategies. Proc. SPIE. 691310–691351

27. Boone J M (1992) Parametrized x-ray absorption in diagnostic radiology from Monte Carlo calculations: implications for x-ray detector design. Med. Phys. 19:1467–1473

28. World Medical Association (2013) World Medical Association Declaration of Helsinki: ethical principles for medical research involving human subjects. JAMA 310:2191–2194

29. Prather H, Hunt D, Steger-May K, Hayes M H, Knaus E, and Clohisy J (2009) Inter-rater reliability of three musculoskeletal physical examination techniques used to assess motion in three planes while standing. PM R 1:629–635

30. Shrout P E and Fleiss J L (1979) Intraclass correlations: uses in assessing rater reliability. Psychol. Bull. 86:420–428

31. Landis J R and Kock G G (1977) The measurement of observer agreement for categorical data. Biometriscs 33:159–174

32. Gupta N and Garg N K (2014) Review on different techniques of image registration. Int. J. Eng. Comput. Sci. ISSN. 2319–7242

33. Wang C-W and Chen H-C (2013) Improved image alignment method in application to X-ray images and biological images. Bioinformatics 29:1879–1887

34. Korsager A S, Carl J, and Ostergaard L R (2016) Comparison of manual and automatic MR-CT registration for radiotherapy of prostate cancer. J. Appl. Clin. Med. Phys. 17:294–303

IMPA

CT O

F PRO

NE

, SUPIN

E A

ND

OB

LIQU

E PA

TIEN

T POSITIO

NIN

G O

N C

BC

T IMA

GE

QU

ALITY

59

4

35. Pauwels R, Araki K, Siewerdsen J H, and Thongvigitmanee S S (2015) Technical aspects of dental CBCT: state of the art. Dentomaxillofacial Radiol. 44:20140224

36. Pereira A M, Batigália F, Santos F S, Martins A T, and Lyra A O (2012) Clinical application of Cone-beam tomography in disorders of the head of mandible. J. Morphol 29:151–153

37. Kwong J C, Palomo J M, Landers M A, Figueroa A, and Hans M G (2008) Image quality produced by different cone-beam computed tomography settings. Am. J. Orthod. Dentofac. Orthop. 133:317–327

38. Sohaib S A, Peppercorn P D, Horrocks J A, Keene M H, Kenyon G S, and Reznek R H (2001) The effect of decreasing mAs on image quality and patient dose in sinus CT. Br. J. Radiol. 74:157–161

39. Timperley D G, Banks C, Robinson D, Roth J, Sacks R, and Harvey R J (2011) Lateral frontal sinus access in endoscopic

skull-base surgery. Int. Forum Allergy Rhinol. 1:290–295

40. Chang C-Y, Ke D-S, and Chen J-Y (2009) Essential fatty acids and human brain. Acta Neurol. Taiwan. 18:231–241

41. Pech P, Bergstrom K, Rauschning W, and Haughton V M (1987) Attenuation values, volume changes and artifacts in tissue due to freezing. Acta Radiol. 28:779–782

42. Hemmingsson A, Johansson A, and Rauschning W (1982) Attenuation in human muscle and fat tissue in vivo and in vitro. Acta Radiol. Diagn. 23:149–151

43. Zou J, Koivisto J, Lähelmä J, Aarnisalo A, Wolff J, and Pyykkö I (2015) Imaging Optimization of Temporal Bones With Cochlear Implant Using a High-resolution Cone Beam CT and the Corresponding Effective Dose. Ann. Otol. Rhinol. Laryngol. 124:466–473

5 CT IMAGE SEGMENTATION METHODS FOR BONE USED IN MEDICAL ADDITIVE MANUFACTURING:

A LITERATURE REVIEW

Maureen van Eijnatten, Roelof van Dijk, Johannes Dobbe,

Geert Streekstra, Juha Koivisto, Jan Wolff

Medical Engineering & Physics [accepted Sept 2017]

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

62

5

ABSTRACTAim of the studyThe accuracy of additive manufactured medical constructs is limited by errors introduced

during image segmentation. The aim of this study was to review the existing literature on

different image segmentation methods used in medical additive manufacturing.

MethodsThirty-two publications that reported on the accuracy of bone segmentation based on

computed tomography images were identified using PubMed, ScienceDirect, Scopus,

and Google Scholar. The advantages and disadvantages of the different segmentation

methods used in these studies were evaluated and reported accuracies were compared.

ResultsThe spread between the reported accuracies was large (0.04 mm to 1.9 mm). Global

thresholding was the most commonly used segmentation method with accuracies under

0.6 mm. The disadvantage of this method is the extensive manual post-processing

required. Advanced thresholding methods could improve the accuracy to under 0.38

mm. However, such methods are currently not included in commercial software packages.

Statistical shape model methods resulted in accuracies from 0.25 mm to 1.9 mm but are

only suitable for anatomical structures with moderate anatomical variations.

ConclusionsThresholding remains the most widely used segmentation method in medical additive

manufacturing. To improve the accuracy and reduce the costs of patient-specific additive

manufactured constructs, more advanced segmentation methods are required.

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

63

5

INTRODUCTIONAdditive manufacturing (AM), also referred to as three-dimensional (3D) printing, is

becoming increasingly popular in medicine [1] since it offers the possibility to personalize

patient care. [2] The use of AM anatomical models results in more precise treatment planning,

better communication, [3,4] and improved training and education. [5,6] Furthermore, AM

can be used for the fabrication of drill guides, [7] saw guides, [8] and patient-specific

implants. [9] To date, medical AM is most commonly used in branches of surgery involving

the musculoskeletal system, such as oral and maxillofacial surgery, traumatology, and

orthopaedic surgery. However, it must be noted that the overall accuracy and repeatability

of medical AM constructs used for bone reconstruction still need to be improved. [10]

In this context, a recent systematic review by Martelli et al. (2016) identified 34 studies

(21.5%) that reported on the unsatisfactory accuracy of medical AM constructs. [11]

The current medical AM process used for the reconstruction of the musculoskeletal

system can be divided into four basic steps: imaging (1); image processing (2), optionally

followed by computer-aided design (3); and additive manufacturing (4). [12] Each of these

steps can introduce geometric deviations that can cause distortions in the resulting medical

AM constructs. [13] Recent studies, however, suggest that the majority of the inaccuracies

are introduced during imaging (Figure 1: step 1) and image processing (Figure 1:

step 2), rather than during the manufacturing, i.e., the 3D printing process, which is

generally considered to be precise. [11,14,15]

Step 1: ImagingCT scanners are best suited for imaging bony structures due to their superior hard tissue

contrast and spatial resolution. [16] Today, a plethora of different CT technologies are

available, ranging from single, helical CT to 128-slice dual-source CT configurations.

Cone-beam computed tomography (CBCT) scanners are becoming increasingly popular in

orthopaedic [17] and maxillofacial surgery [18] due to their lower radiation dose and costs.

Raw CT data acquired during image acquisition is commonly reconstructed as a Digital

Imaging and Communications in Medicine (DICOM) file.

One major challenge faced in medical AM is the large variety of different CT image

acquisition and reconstruction parameters currently available (see Figure 1; step 1 A&B).

To date, to the best of our knowledge, there are no standardized protocols available for

medical AM. Image slice thickness and slice interval have been identified as the primary

limiting factors for the overall accuracy of medical AM constructs, [19] especially when

reconstructing thin bony structures from axial plane images, such as the orbital floor, [20]

or where the imaging plane is nearly parallel to the bone surface to be reconstructed, such

as in the tibial plateau. Moreover, imaging noise, beam hardening, patient motion, and

metal artifacts can introduce inhomogeneities in the gray values of CT images.

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

64

5

Step 2: Image processingThe medical AM process always requires image processing: the conversion of CT

images into 3D surface models (Figure 1, step 2). Such 3D surface models can be saved

in a wide range of different file formats. [21] Currently, the most commonly used file

format in medical AM is standard tessellation language (STL). Although rarely used, it is

theoretically possible to convert CT images into slice formats that can subsequently be

used for AM, thereby skipping the STL conversion process. However, it must be noted that

the computer-aided design (CAD) software packages currently available on the market for

medical AM still require STL files (Figure 1, step 3).

The conversion of DICOM datasets into 3D surface models has been identified as

a major source of inaccuracies in medical AM. [14,22,23] The most important step in this

conversion process is image segmentation (Figure 1, step 2A). [24] Segmentation refers

to the partitioning of images into regions of interest (ROIs) that correspond to anatomical

structures. At present, there are a plethora of different image segmentation methods

available for bony structures. [25] It remains unclear, however, which segmentation method

offers the most accurate 3D surface models. Therefore, the first aim of this study was to

review the existing literature on the different CT image segmentation methods currently

being used for bone segmentation in medical AM. The second aim was to evaluate

the impact of the different image segmentation methods on the geometric accuracy of

3D surface models.

MATERIALS AND METHODSExisting literature on the CT image segmentation of bony structures for medical AM

applications was reviewed using the PubMed Medline literature database, ScienceDirect,

Scopus, and Google Scholar. An initial database of 17,700 publications was generated

using the search terms “medical AND X” + “medicine AND X”, with X representing

the interchangeably used terms “additive manufacturing”, “rapid prototyping”, and “3D

printing”. The acquired database was subsequently filtered using the search terms “bone”

(7,630), “segmentation” (281), and “accuracy” (138). In order to acquire comparable

results, the following inclusion and exclusion criteria were used to isolate the publications

of interest to this study:

Inclusion criteria:

• Evaluates the geometric accuracy of 3D surface models of bone derived from CT

images

• Reports accuracy using a mean spatial distance-based metric in millimeters

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

65

5

Exclusion criteria:

• Uses non-biological materials or artificial phantoms

• Reports accuracy using non-spatial distance-based metrics, such as relative distance or

volume difference (%), maximum distance, 95%-value of all distances (95th percentile),

overlap-based metrics, and contour-based metrics.

• Uses a non-CT-based imaging technology, e.g., magnetic resonance imaging (MRI)

• Only reports on the accuracy of the final AM model and not on the preceding 3D

surface model

Figure 1. Overview of the parameters that can influence the accuracy of medical AM constructs. * This review focuses on the different CT image segmentation methods used in medical AM.

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

66

5

Using the aforementioned inclusion and exclusion criteria, 10 publications were

identified that contained information on the geometric accuracy of 3D surface models

of bone acquired from CT images that were subsequently used for medical AM. From

a parallel search, 22 additional publications were identified that reported on the accuracy

of 3D surface models but did not focus on AM applications. Additional information was

collected from four review articles focusing on medical AM. [19,26–28]

All relevant details of the publications included in this study are summarized in

Table 1. The publications were subsequently sorted based on the type of image

segmentation method used in the study. The reported accuracies and standard deviations

(if available) are summarized in Table 2. In cases where multiple image segmentation

methods, software packages, anatomical structures, or CT modalities were evaluated, their

recorded accuracies are presented separately in Table 2. When multiple measurements

were performed using the same image segmentation method or anatomical structure,

the mean value and standard deviation of the reported accuracies are presented as one

entry in Table 2.

RESULTSA total of 32 publications were included in this study (Table 1). These publications

were published between 2002 and 2017. To the best of our knowledge, no papers on

the accuracy of medical AM constructs were published before 2002. [29] However, it must

be noted that the first publications on medical AM date back as far as the early 1990s. [30]

CT image segmentation methods used in medical AMThe reviewed publications reported that the following three semi-automated CT image

segmentation methods are often used in medical AM: global thresholding, [31] edge

detection, [32] and region growing. [33]

Of the image segmentation methods mentioned above, global thresholding is the most

commonly used bone segmentation method in medical AM. In global thresholding,

a single threshold value t for bone is manually [34] or automatically [31,35] selected. All

voxels with a gray value equal or greater than t are included in a segmented volume using

a binary mask Mx,y (Equation 1):

(1)

where Ix,y denotes the grey value at coordinates x and y in a CT image slice.

Global thresholding has certain drawbacks, namely, voxels reside on tissue boundaries

that contain more than one tissue type and induce a blurring of gray values across

boundaries. This phenomenon is referred to as the partial volume effect (PVE). As

 

8

Results A total of 32 publications were included in this study (Table 1). These publications were published

between 2002 and 2017. To the best of our knowledge, no papers on the accuracy of medical AM

constructs were published before 2002. [29] However, it must be noted that the first publications

on medical AM date back as far as the early 1990s. [30]

CT image segmentation methods used in medical AM

The reviewed publications reported that the following three semi-automated CT image

segmentation methods are often used in medical AM: global thresholding, [31] edge detection, [32]

and region growing. [33]

Of the image segmentation methods mentioned above, global thresholding is the most commonly

used bone segmentation method in medical AM. In global thresholding, a single threshold value t

for bone is manually [34] or automatically [31,35] selected. All voxels with a gray value equal or

greater than t are included in a segmented volume using a binary mask Mx,y (Equation 1):

M�,� � �1I�,� � �0I�,� � �,�1�

where Ix,y denotes the grey value at coordinates x and y in a CT image slice.

Global thresholding has certain drawbacks, namely, voxels reside on tissue boundaries that contain

more than one tissue type and induce a blurring of gray values across boundaries. This phenomenon

is referred to as the partial volume effect (PVE). As a consequence of the PVE, precise delineation

of tissue boundaries using only a single threshold value remains difficult and can result in an over-

or underestimation of the region of interest. Furthermore, a single threshold value does not take

artifacts, image noise, photon starvation, or variations in gray values between different CT scanners

and protocols into account, which can lead to inconsistent segmentation results. These drawbacks

have led to the development of more sophisticated thresholding methods such as local thresholding.

Local thresholding, also referred to as multi-level thresholding, divides an image into multiple

ROIs for which an individual thresholding bandwidth tk to tk+1 can be selected. [36] All voxels with

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

67

5

a consequence of the PVE, precise delineation of tissue boundaries using only a single

threshold value remains difficult and can result in an over- or underestimation of the region

of interest. Furthermore, a single threshold value does not take artifacts, image noise,

photon starvation, or variations in gray values between different CT scanners and protocols

into account, which can lead to inconsistent segmentation results. These drawbacks

have led to the development of more sophisticated thresholding methods such as

local thresholding.

Local thresholding, also referred to as multi-level thresholding, divides an image into

multiple ROIs for which an individual thresholding bandwidth tk to tk+1 can be selected. [36]

All voxels with a grey value between tk and tk+1 are included in k segmented volumes using

a binary mask Mx,y (Equation 2):

(2)

where k denotes the kth grey value band, of which tk is the lower limit and tk+1 the upper

limit. Furthermore, 3D adaptive thresholding algorithms have recently been developed

that offer the possibility to iteratively update the voxel classification in an additional

second step after global thresholding. [37]

Edge detection, on the other hand, identifies local edges on CT images by calculating

gray value gradients (derivatives). The gradients with a magnitude that is higher than

a chosen threshold value are defined as edges. A range of edge detection operators are

currently available that includes Sobel, Laplacian (2nd derivative) and Canny. [25] Canny edge

detection remains one of the most commonly used, fastest and most accurate operators.

[36] Edge detection is well suited for the segmentation of structures with different contrast

in different regions, such as long bones. However, local gray value changes induced by

noise and (metal) artifacts are often mistakenly identified as edges. It should also be noted

that, without further processing, edge detection methods do not necessarily segment all

bone voxels in the image and thus need to be combined with other methods, such as

region growing.

In region growing, a specific voxel is manually selected as a seed point for a specific

tissue type. Subsequently, the gray values of the neighboring voxels are compared to

the gray value of the seed point. Voxels that meet the predefined homogeneity criteria are

labeled and grouped together, thereby creating a region Ri of connected voxels in the CT

image I:

 

9

a grey value between tk and tk+1 are included in k segmented volumes using a binary mask Mx,y

(Equation 2):

M�,� �

��������,� � ����� � ��,� � ����� � ��,� � ����� � ��,� � �������,� � ����

,���

where k denotes the kth grey value band, of which tk is the lower limit and tk+1 the upper limit.

Furthermore, 3D adaptive thresholding algorithms have recently been developed that offer the

possibility to iteratively update the voxel classification in an additional second step after global

thresholding. [37]

Edge detection, on the other hand, identifies local edges on CT images by calculating gray value

gradients (derivatives). The gradients with a magnitude that is higher than a chosen threshold value

are defined as edges. A range of edge detection operators are currently available that includes

Sobel, Laplacian (2nd derivative) and Canny. [25] Canny edge detection remains one of the most

commonly used, fastest and most accurate operators. [36] Edge detection is well suited for the

segmentation of structures with different contrast in different regions, such as long bones.

However, local gray value changes induced by noise and (metal) artifacts are often mistakenly

identified as edges. It should also be noted that, without further processing, edge detection methods

do not necessarily segment all bone voxels in the image and thus need to be combined with other

methods, such as region growing.

In region growing, a specific voxel is manually selected as a seed point for a specific tissue type.

Subsequently, the gray values of the neighboring voxels are compared to the gray value of the seed

point. Voxels that meet the predefined homogeneity criteria are labeled and grouped together,

thereby creating a region Ri of connected voxels in the CT image I:

�� � �� � �� � �� �� � ��3�

In practice, region growing is seldom the only segmentation method used, but it is commonly

combined with other methods such as (global) thresholding. An advantage of region growing is

that it discards voxels that are not connected to the anatomical structure of interest, resulting in a

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

68

5

(3)

In practice, region growing is seldom the only segmentation method used, but it is

commonly combined with other methods such as (global) thresholding. An advantage of

region growing is that it discards voxels that are not connected to the anatomical structure

of interest, resulting in a shorter 3D printing time and material savings. A disadvantage

of region growing is that each separate bony structure requires an individual, manually

placed seed point. Furthermore, noise and the partial volume effect can cause voids or

erroneously connected structures in a segmented image.

In recent years, a number of advanced CT image segmentation methods have been

developed. [38] The most established and validated advanced image segmentation

method to date is based on statistical shape models (SSM). SSM-based methods generate

a statistical shape model of an anatomical ROI from a training dataset. This training

dataset comprises a large number of fully segmented CT images, which are averaged to

form a mean anatomical shape model, along with common modes of variation and their

probabilities. The model given by:

(4)

describes every possible shape

 

10

shorter 3D printing time and material savings. A disadvantage of region growing is that each

separate bony structure requires an individual, manually placed seed point. Furthermore, noise and

the partial volume effect can cause voids or erroneously connected structures in a segmented image.

In recent years, a number of advanced CT image segmentation methods have been developed. [38]

The most established and validated advanced image segmentation method to date is based on

statistical shape models (SSM). SSM-based methods generate a statistical shape model of an

anatomical ROI from a training dataset. This training dataset comprises a large number of fully

segmented CT images, which are averaged to form a mean anatomical shape model, along with

common modes of variation and their probabilities. The model given by:

��������� � �� � ∑ ��� ��, (4)

describes every possible shape �� as a linear combination of k eigenmodes �� with weights ���, where

�� denotes the mean shape model. This model can be fitted to a new, unsegmented image dataset,

and the segmentation in terms of weights ��� then forms an approximation of the segmented

shape.[39]

The major challenge faced in SSM-based segmentation is the large number of uniform training

datasets required. Another drawback with SSM-based methods is that they are prone to failure if

the datasets differ strongly from the training shapes. In clinical settings, such differences can occur,

for example, in fractured bony structures that no longer correspond to the initial anatomical shape.

Another example is that an SSM trained on a European population may not account for some of

the morphological differences present in an African population.

as a linear combination of k eigenmodes

 

10

shorter 3D printing time and material savings. A disadvantage of region growing is that each

separate bony structure requires an individual, manually placed seed point. Furthermore, noise and

the partial volume effect can cause voids or erroneously connected structures in a segmented image.

In recent years, a number of advanced CT image segmentation methods have been developed. [38]

The most established and validated advanced image segmentation method to date is based on

statistical shape models (SSM). SSM-based methods generate a statistical shape model of an

anatomical ROI from a training dataset. This training dataset comprises a large number of fully

segmented CT images, which are averaged to form a mean anatomical shape model, along with

common modes of variation and their probabilities. The model given by:

��������� � �� � ∑ ��� ��, (4)

describes every possible shape �� as a linear combination of k eigenmodes �� with weights ���, where

�� denotes the mean shape model. This model can be fitted to a new, unsegmented image dataset,

and the segmentation in terms of weights ��� then forms an approximation of the segmented

shape.[39]

The major challenge faced in SSM-based segmentation is the large number of uniform training

datasets required. Another drawback with SSM-based methods is that they are prone to failure if

the datasets differ strongly from the training shapes. In clinical settings, such differences can occur,

for example, in fractured bony structures that no longer correspond to the initial anatomical shape.

Another example is that an SSM trained on a European population may not account for some of

the morphological differences present in an African population.

with

weights

 

10

shorter 3D printing time and material savings. A disadvantage of region growing is that each

separate bony structure requires an individual, manually placed seed point. Furthermore, noise and

the partial volume effect can cause voids or erroneously connected structures in a segmented image.

In recent years, a number of advanced CT image segmentation methods have been developed. [38]

The most established and validated advanced image segmentation method to date is based on

statistical shape models (SSM). SSM-based methods generate a statistical shape model of an

anatomical ROI from a training dataset. This training dataset comprises a large number of fully

segmented CT images, which are averaged to form a mean anatomical shape model, along with

common modes of variation and their probabilities. The model given by:

��������� � �� � ∑ ��� ��, (4)

describes every possible shape �� as a linear combination of k eigenmodes �� with weights ���, where

�� denotes the mean shape model. This model can be fitted to a new, unsegmented image dataset,

and the segmentation in terms of weights ��� then forms an approximation of the segmented

shape.[39]

The major challenge faced in SSM-based segmentation is the large number of uniform training

datasets required. Another drawback with SSM-based methods is that they are prone to failure if

the datasets differ strongly from the training shapes. In clinical settings, such differences can occur,

for example, in fractured bony structures that no longer correspond to the initial anatomical shape.

Another example is that an SSM trained on a European population may not account for some of

the morphological differences present in an African population.

, where

 

10

shorter 3D printing time and material savings. A disadvantage of region growing is that each

separate bony structure requires an individual, manually placed seed point. Furthermore, noise and

the partial volume effect can cause voids or erroneously connected structures in a segmented image.

In recent years, a number of advanced CT image segmentation methods have been developed. [38]

The most established and validated advanced image segmentation method to date is based on

statistical shape models (SSM). SSM-based methods generate a statistical shape model of an

anatomical ROI from a training dataset. This training dataset comprises a large number of fully

segmented CT images, which are averaged to form a mean anatomical shape model, along with

common modes of variation and their probabilities. The model given by:

��������� � �� � ∑ ��� ��, (4)

describes every possible shape �� as a linear combination of k eigenmodes �� with weights ���, where

�� denotes the mean shape model. This model can be fitted to a new, unsegmented image dataset,

and the segmentation in terms of weights ��� then forms an approximation of the segmented

shape.[39]

The major challenge faced in SSM-based segmentation is the large number of uniform training

datasets required. Another drawback with SSM-based methods is that they are prone to failure if

the datasets differ strongly from the training shapes. In clinical settings, such differences can occur,

for example, in fractured bony structures that no longer correspond to the initial anatomical shape.

Another example is that an SSM trained on a European population may not account for some of

the morphological differences present in an African population.

denotes the mean shape model. This model can be fitted to a new,

unsegmented image dataset, and the segmentation in terms of weights

 

10

shorter 3D printing time and material savings. A disadvantage of region growing is that each

separate bony structure requires an individual, manually placed seed point. Furthermore, noise and

the partial volume effect can cause voids or erroneously connected structures in a segmented image.

In recent years, a number of advanced CT image segmentation methods have been developed. [38]

The most established and validated advanced image segmentation method to date is based on

statistical shape models (SSM). SSM-based methods generate a statistical shape model of an

anatomical ROI from a training dataset. This training dataset comprises a large number of fully

segmented CT images, which are averaged to form a mean anatomical shape model, along with

common modes of variation and their probabilities. The model given by:

��������� � �� � ∑ ��� ��, (4)

describes every possible shape �� as a linear combination of k eigenmodes �� with weights ���, where

�� denotes the mean shape model. This model can be fitted to a new, unsegmented image dataset,

and the segmentation in terms of weights ��� then forms an approximation of the segmented

shape.[39]

The major challenge faced in SSM-based segmentation is the large number of uniform training

datasets required. Another drawback with SSM-based methods is that they are prone to failure if

the datasets differ strongly from the training shapes. In clinical settings, such differences can occur,

for example, in fractured bony structures that no longer correspond to the initial anatomical shape.

Another example is that an SSM trained on a European population may not account for some of

the morphological differences present in an African population.

then forms an

approximation of the segmented shape.[39]

The major challenge faced in SSM-based segmentation is the large number of

uniform training datasets required. Another drawback with SSM-based methods is that

they are prone to failure if the datasets differ strongly from the training shapes. In clinical

settings, such differences can occur, for example, in fractured bony structures that no

longer correspond to the initial anatomical shape. Another example is that an SSM trained

on a European population may not account for some of the morphological differences

present in an African population.

Accuracy of CT image segmentation methodsIn all the reviewed publications, the accuracy was assessed by comparing the CT-derived

3D surface models to a “ground truth” (Table 1). The ground truth was obtained either by

performing manual segmentation by an expert, by obtaining a laser surface scan or optical

3D surface scan from the dry or dissected bone, or by performing linear measurements on

the dry or dissected bone using calipers or a coordinate measurement machine (CMM).

In all publications that focused on SSM-based segmentation methods, the accuracy was

 

9

a grey value between tk and tk+1 are included in k segmented volumes using a binary mask Mx,y

(Equation 2):

M�,� �

��������,� � ����� � ��,� � ����� � ��,� � ����� � ��,� � �������,� � ����

,���

where k denotes the kth grey value band, of which tk is the lower limit and tk+1 the upper limit.

Furthermore, 3D adaptive thresholding algorithms have recently been developed that offer the

possibility to iteratively update the voxel classification in an additional second step after global

thresholding. [37]

Edge detection, on the other hand, identifies local edges on CT images by calculating gray value

gradients (derivatives). The gradients with a magnitude that is higher than a chosen threshold value

are defined as edges. A range of edge detection operators are currently available that includes

Sobel, Laplacian (2nd derivative) and Canny. [25] Canny edge detection remains one of the most

commonly used, fastest and most accurate operators. [36] Edge detection is well suited for the

segmentation of structures with different contrast in different regions, such as long bones.

However, local gray value changes induced by noise and (metal) artifacts are often mistakenly

identified as edges. It should also be noted that, without further processing, edge detection methods

do not necessarily segment all bone voxels in the image and thus need to be combined with other

methods, such as region growing.

In region growing, a specific voxel is manually selected as a seed point for a specific tissue type.

Subsequently, the gray values of the neighboring voxels are compared to the gray value of the seed

point. Voxels that meet the predefined homogeneity criteria are labeled and grouped together,

thereby creating a region Ri of connected voxels in the CT image I:

�� � �� � �� � �� �� � ��3�

In practice, region growing is seldom the only segmentation method used, but it is commonly

combined with other methods such as (global) thresholding. An advantage of region growing is

that it discards voxels that are not connected to the anatomical structure of interest, resulting in a

 

10

shorter 3D printing time and material savings. A disadvantage of region growing is that each

separate bony structure requires an individual, manually placed seed point. Furthermore, noise and

the partial volume effect can cause voids or erroneously connected structures in a segmented image.

In recent years, a number of advanced CT image segmentation methods have been developed. [38]

The most established and validated advanced image segmentation method to date is based on

statistical shape models (SSM). SSM-based methods generate a statistical shape model of an

anatomical ROI from a training dataset. This training dataset comprises a large number of fully

segmented CT images, which are averaged to form a mean anatomical shape model, along with

common modes of variation and their probabilities. The model given by:

��������� � �� � ∑ ��� ��, (4)

describes every possible shape �� as a linear combination of k eigenmodes �� with weights ���, where

�� denotes the mean shape model. This model can be fitted to a new, unsegmented image dataset,

and the segmentation in terms of weights ��� then forms an approximation of the segmented

shape.[39]

The major challenge faced in SSM-based segmentation is the large number of uniform training

datasets required. Another drawback with SSM-based methods is that they are prone to failure if

the datasets differ strongly from the training shapes. In clinical settings, such differences can occur,

for example, in fractured bony structures that no longer correspond to the initial anatomical shape.

Another example is that an SSM trained on a European population may not account for some of

the morphological differences present in an African population.

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

69

5

evaluated using leave-one-out cross validation, in which the distances were measured

from the omitted training shape to the closest match of the “ground-truth” shape model

trained on all other available shapes.

In total, 22 of the reviewed publications (67%) evaluated the accuracy by calculating

the mean absolute distance between a CT-derived 3D surface model and a geometrically

aligned ground truth 3D surface model. Let us define the distance d(p,q) as the distance

between a point p belonging to the surface of a 3D surface model P and its closest point q

on the ground truth model Q. The mean surface distance (MSD) between the two surfaces

P and Q can then be defined as follows:

(5)

where |P| denotes the surface area of P. [40] Note that some variations in the definition

of d(p,q) were found in the reviewed publications. For example, Smith et al. (2013) used

perpendicular projections of the distance of each point p along the triangle normal vector,

[14] whilst Gassman et al. (2008) created a distance map based on the Euclidean distance,

or the shortest distance from each point p to the target surface Q. [41] Finally, Zhang et

al. [42] and Rajamani et al. [43] used the Hausdorff distance (HD), which can be defined

as follows:

(6)

Two publications [44,45] used the root-mean-square distance (RMSD) that can be derived

from Equation 5:

(7)

In four publications, [29,46–48] the mean absolute distance (MAD) was calculated by

performing a limited number of linear measurements n, which can be written as follows:

(8)

 

11

Accuracy of CT image segmentation methods In all the reviewed publications, the accuracy was assessed by comparing the CT-derived 3D

surface models to a “ground truth” (Table 1). The ground truth was obtained either by performing

manual segmentation by an expert, by obtaining a laser surface scan or optical 3D surface scan

from the dry or dissected bone, or by performing linear measurements on the dry or dissected bone

using calipers or a coordinate measurement machine (CMM). In all publications that focused on

SSM-based segmentation methods, the accuracy was evaluated using leave-one-out cross

validation, in which the distances were measured from the omitted training shape to the closest

match of the “ground-truth” shape model trained on all other available shapes.

In total, 22 of the reviewed publications (67%) evaluated the accuracy by calculating the mean

absolute distance between a CT-derived 3D surface model and a geometrically aligned ground truth

3D surface model. Let us define the distance d(p,q) as the distance between a point p belonging to

the surface of a 3D surface model P and its closest point q on the ground truth model Q. The mean

surface distance (MSD) between the two surfaces P and Q can then be defined as follows:

��� � 1|�|� ���, �������

,�5�

where |P| denotes the surface area of P. [40] Note that some variations in the definition of d(p,q)

were found in the reviewed publications. For example, Smith et al. (2013) used perpendicular

projections of the distance of each point p along the triangle normal vector, [14] whilst Gassman

et al. (2008) created a distance map based on the Euclidean distance, or the shortest distance from

each point p to the target surface Q. [41] Finally, Zhang et al. [42] and Rajamani et al. [43] used

the Hausdorff distance (HD), which can be defined as follows:

����, �� � ������ ������� ����, �������

Two publications [44,45] used the root-mean-square distance (RMSD) that can be derived from

Equation 5:

���� � � 1|�|� ���, �����

����7�

 

11

Accuracy of CT image segmentation methods In all the reviewed publications, the accuracy was assessed by comparing the CT-derived 3D

surface models to a “ground truth” (Table 1). The ground truth was obtained either by performing

manual segmentation by an expert, by obtaining a laser surface scan or optical 3D surface scan

from the dry or dissected bone, or by performing linear measurements on the dry or dissected bone

using calipers or a coordinate measurement machine (CMM). In all publications that focused on

SSM-based segmentation methods, the accuracy was evaluated using leave-one-out cross

validation, in which the distances were measured from the omitted training shape to the closest

match of the “ground-truth” shape model trained on all other available shapes.

In total, 22 of the reviewed publications (67%) evaluated the accuracy by calculating the mean

absolute distance between a CT-derived 3D surface model and a geometrically aligned ground truth

3D surface model. Let us define the distance d(p,q) as the distance between a point p belonging to

the surface of a 3D surface model P and its closest point q on the ground truth model Q. The mean

surface distance (MSD) between the two surfaces P and Q can then be defined as follows:

��� � 1|�|� ���, �������

,�5�

where |P| denotes the surface area of P. [40] Note that some variations in the definition of d(p,q)

were found in the reviewed publications. For example, Smith et al. (2013) used perpendicular

projections of the distance of each point p along the triangle normal vector, [14] whilst Gassman

et al. (2008) created a distance map based on the Euclidean distance, or the shortest distance from

each point p to the target surface Q. [41] Finally, Zhang et al. [42] and Rajamani et al. [43] used

the Hausdorff distance (HD), which can be defined as follows:

����, �� � ������ ������� ����, �������

Two publications [44,45] used the root-mean-square distance (RMSD) that can be derived from

Equation 5:

���� � � 1|�|� ���, �����

����7�

 

11

Accuracy of CT image segmentation methods In all the reviewed publications, the accuracy was assessed by comparing the CT-derived 3D

surface models to a “ground truth” (Table 1). The ground truth was obtained either by performing

manual segmentation by an expert, by obtaining a laser surface scan or optical 3D surface scan

from the dry or dissected bone, or by performing linear measurements on the dry or dissected bone

using calipers or a coordinate measurement machine (CMM). In all publications that focused on

SSM-based segmentation methods, the accuracy was evaluated using leave-one-out cross

validation, in which the distances were measured from the omitted training shape to the closest

match of the “ground-truth” shape model trained on all other available shapes.

In total, 22 of the reviewed publications (67%) evaluated the accuracy by calculating the mean

absolute distance between a CT-derived 3D surface model and a geometrically aligned ground truth

3D surface model. Let us define the distance d(p,q) as the distance between a point p belonging to

the surface of a 3D surface model P and its closest point q on the ground truth model Q. The mean

surface distance (MSD) between the two surfaces P and Q can then be defined as follows:

��� � 1|�|� ���, �������

,�5�

where |P| denotes the surface area of P. [40] Note that some variations in the definition of d(p,q)

were found in the reviewed publications. For example, Smith et al. (2013) used perpendicular

projections of the distance of each point p along the triangle normal vector, [14] whilst Gassman

et al. (2008) created a distance map based on the Euclidean distance, or the shortest distance from

each point p to the target surface Q. [41] Finally, Zhang et al. [42] and Rajamani et al. [43] used

the Hausdorff distance (HD), which can be defined as follows:

����, �� � ������ ������� ����, �������

Two publications [44,45] used the root-mean-square distance (RMSD) that can be derived from

Equation 5:

���� � � 1|�|� ���, �����

����7�

 

12

In four publications, [29,46–48] the mean absolute distance (MAD) was calculated by performing

a limited number of linear measurements n, which can be written as follows:

��� � 1��|�� � ��|

�,���

where D is a linear measurement performed on the ground truth, and x is a measurement performed

on the CT-derived 3D surface model. Additionally, in four publications, [49–52] the mean distance

(MD) was calculated as the signed mean of n measurements between different anatomical

landmarks, [49,52] reference points, [51] or photographs of a cryosectioned cadaver, [50] which

can be described as follows:

�� � 1����� � ���

����

Table 2 summarizes the geometric accuracies reported in the publications. All 3D surface model

accuracies generally varied between 0.04 mm and 1.9 mm. Manual segmentation resulted in an

accuracy of 0.20 mm ± 0.15 mm to 0.48 mm ± 0.51 mm. [53,54] In global thresholding, accuracies

varied between 0.04 mm ± 0.59 mm and 0.62 mm ± 0.76 mm. [14,29,36,46,48,49,52,55–59]

However, it must be noted that most authors reported that additional manual editing was performed.

Advanced thresholding methods resulted in accuracies below 0.38 mm ± 0.14 mm.

[36,37,42,45,46] SSM-based segmentation methods achieved accuracies ranging from 0.25 mm ±

0.20 mm to 1.9 mm ± 0.20 mm. [44,60–68] Finally, the publications on more experimental

segmentation methods (“other” in Table 2) reported geometric accuracies ranging between 0.28

mm ± 0.04 mm and 1.27 mm ± 0.92 mm. [41,47,50,51,59,69,70]

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

70

5

where D is a linear measurement performed on the ground truth, and x is a measurement

performed on the CT-derived 3D surface model. Additionally, in four publications,

[49–52] the mean distance (MD) was calculated as the signed mean of n measurements

between different anatomical landmarks, [49,52] reference points, [51] or photographs of

a cryosectioned cadaver, [50] which can be described as follows:

(9)

Table 2 summarizes the geometric accuracies reported in the publications. All 3D surface

model accuracies generally varied between 0.04 mm and 1.9 mm. Manual segmentation

resulted in an accuracy of 0.20 mm ± 0.15 mm to 0.48 mm ± 0.51 mm. [53,54] In global

thresholding, accuracies varied between 0.04 mm ± 0.59 mm and 0.62 mm ± 0.76 mm.

[14,29,36,46,48,49,52,55–59] However, it must be noted that most authors reported that

additional manual editing was performed. Advanced thresholding methods resulted in

accuracies below 0.38 mm ± 0.14 mm. [36,37,42,45,46] SSM-based segmentation methods

achieved accuracies ranging from 0.25 mm ± 0.20 mm to 1.9 mm ± 0.20 mm. [44,60–68]

Finally, the publications on more experimental segmentation methods (“other” in Table 2)

reported geometric accuracies ranging between 0.28 mm ± 0.04 mm and 1.27 mm ± 0.92

mm. [41,47,50,51,59,69,70]

DISCUSSIONTo date, the most common CT image segmentation method used for bone segmentation

in medical AM is global thresholding. All publications reviewed in this study that used

global thresholding reported geometric accuracies under 0.62 mm ± 0.76 mm [59]

(Table 2). However, the reported accuracies generally included additional extensive manual

post-processing, which is often very time consuming. Furthermore, the reproducibility of

manual post-processing remains a challenge, especially when suboptimal CT images with

thick slices, noise, metal artifacts or motion artifacts are used. In this context, Fasel et al.

[56] reported that a manual clean-up time of 10 hours was required to achieve an accuracy

of 0.25 mm ± 0.23 mm. Such extensive post-processing times are currently a major reason

for the high costs related to medical AM. [27]

In order to overcome this time-consuming manual task, more advanced thresholding

methods can be applied. In this context, an algorithm using multiple local thresholds,

automatically determined using edge detection, achieved an accuracy of 0.18 mm ± 0.02

mm without the need for any supplementary manual editing. [36] Furthermore, another

advanced thresholding method that iteratively reclassifies boundary voxels according to

their surroundings resulted in accuracies of 0.08 mm (calcaneus) and 0.20 mm (spine).

[37,42] However, to the best of our knowledge, such advanced thresholding methods

 

12

In four publications, [29,46–48] the mean absolute distance (MAD) was calculated by performing

a limited number of linear measurements n, which can be written as follows:

��� � 1��|�� � ��|

�,���

where D is a linear measurement performed on the ground truth, and x is a measurement performed

on the CT-derived 3D surface model. Additionally, in four publications, [49–52] the mean distance

(MD) was calculated as the signed mean of n measurements between different anatomical

landmarks, [49,52] reference points, [51] or photographs of a cryosectioned cadaver, [50] which

can be described as follows:

�� � 1����� � ���

����

Table 2 summarizes the geometric accuracies reported in the publications. All 3D surface model

accuracies generally varied between 0.04 mm and 1.9 mm. Manual segmentation resulted in an

accuracy of 0.20 mm ± 0.15 mm to 0.48 mm ± 0.51 mm. [53,54] In global thresholding, accuracies

varied between 0.04 mm ± 0.59 mm and 0.62 mm ± 0.76 mm. [14,29,36,46,48,49,52,55–59]

However, it must be noted that most authors reported that additional manual editing was performed.

Advanced thresholding methods resulted in accuracies below 0.38 mm ± 0.14 mm.

[36,37,42,45,46] SSM-based segmentation methods achieved accuracies ranging from 0.25 mm ±

0.20 mm to 1.9 mm ± 0.20 mm. [44,60–68] Finally, the publications on more experimental

segmentation methods (“other” in Table 2) reported geometric accuracies ranging between 0.28

mm ± 0.04 mm and 1.27 mm ± 0.92 mm. [41,47,50,51,59,69,70]

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

71

5

are currently not implemented in commercially available medical AM software packages.

In this context, Kang et al. [49] compared four different commercial software packages

and reported that global thresholding was the segmentation method of choice in all four

software packages.

Currently, only two commercial software packages for medical AM have been approved

by the United States Food and Drug Administration (FDA): MimicsTM (Materialise®,

Belgium), and Osirix® MD (Pixmeo, Switzerland). [71] Furthermore D2P™ (DICOM TO

PRINT) (3D Systems, Rock Hill, USA) has very recently (Jan 9, 2017) received 510k clearance

and is currently pending CE approval. These software packages are generally equipped

with global thresholding, region growing, dynamic region growing, local editing by a local

threshold, and multiple manual mask editing tools (e.g., delete, draw). More advanced

semi-automated image segmentation algorithms are currently available in Amira® 3D

analysis software for Life Sciences (FEI Visualization Sciences Group, Hillsboro, OR, USA)

that allow for interactive segmentation using local image histograms and edge detection

algorithms. However, this software is not presently cleared for medical use, although this

will be a requirement from 2020 onwards under the new European Union medical device

regulation (MDR) for patient-specific constructs. [72] In this context, standardization and

CE certification will be required for the whole medical AM process (Figure 1), including

the evaluation of the repeatability and accuracy of the image segmentation process.

The previously discussed drawbacks of global and advanced thresholding have

led to the development of novel SSM-based CT image segmentation approaches.

However, the SSM publications assessed in this study demonstrated a large variability in

segmentation accuracies (Table 2). Moreover, only two out of eleven publications (18%)

reported accuracies equal to or less than 0.5 mm. [60,61] The major challenge in SSM-based

segmentation methods is the quality and number of available training datasets, which

can affect the generalization ability and specificity of an SSM. [39] The generalization

ability quantifies to which extent the variability of the anatomical shapes in the patient

population are incorporated in the SSM. For example, when too few training datasets

are used, the SSM may fail to represent new shapes. The specificity of an SSM refers to

its ability to only generate segmented shapes that are similar to those incorporated in

the training datasets. Therefore, SSM-based segmentation methods are only applicable

for CT datasets with moderate anatomical variations. Note that 3D statistical deformable

models have also been used to segment bony structures, such as the knee joint, on

magnetic resonance images. [73]

Clinical implicationsThe majority of the accuracies that are depicted in Table 2 may be considered sufficient

for the creation of anatomical models, [2] of which hundreds are being successfully

manufactured and used in clinical settings every year. [74,75] However, the use of inaccurate

3D surface models to create patient-specific AM implants or guides can lead to ill-fitting

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

72

5

Table 1. Publications included in this review.

Ref Year Author Segmentation method Software Anatomy (#) MaterialCT scanner (manufacturer/# slices)

CT tube voltage / tube current

CT slice thickness / voxel size Ground truth

55 2016 Szymor Global thresholding Slicer Skull (1) Dry bone CBCT (Accuitomo) 90 kV / 5 mA 0.5 mm / 0.25 mm

Optical 3D scan

61 2015 Chu Random forest & multi-atlas & articulated SSM

Custom & Amira Femur (60) + acet. (60)

In-vivo MDCT 1.6 mm / 0.6 mm

Manual segmentation

48 2015 Poleti Global thresholding Dolphin (A)/ InVesalius (B)

Mandible (10) Dry bone CBCT (i-CAT) 120 kV / 18.5 mAs - / 0.3 mm Digital calipers

64 2014 Balestra Articulated SSM Custom & Amira Femur (26) + pelvis (26)

In-vivo MDCT Manual segmentation

56 2014 Fasel Global thresholding n.a. (3 institutes) Skull (3) Cadaver MDCT (Siemens Sensation/64) Optical 3D scan49 2014 Kang Global thresholding InVivoDental (A)

/ Mimics (B) / OnDemand3D (C) / OsiriX (D)

Skull (3) Dry bone MDCT (Siemens Sensation/64) 0.6 mm / 0.44 mm

Optical 3D scan

46 2014 Santolaria Global thresholding (A) / Advanced thresholding (B)

Mimics Mandible (1) Dry bone CBCT (Kodak 9500) - / 0.2 mm - 0.3 mm

CMM

57 2014 Van den Broeck Global thresholding Mimics Tibia (10) Cadaver MDCT (GE LightSpeed VCT) 120 kVp / 160 mAs

0.63 mm / 0.39 mm

Optical 3D scan

44 2014 Zhang Region-based SSM Custom & Stradwin Femur (41) In-vivo MDCT 1.6 mm / 1.0 mm

Manual segmentation

60 2013 Chang Base invariant wavelet active shape model

Custom Maxilla (19) In-vivo CBCT (Sirona) - / 0.29 mm Manual segmentation

47 2013 Engelbrecht “Commercial company” (A) / “doctor” (B)

Mimics & SimPlant Ortho Pro

Mandible (7) Cadaver CBCT (KaVo 3D exam) - / 0.3 mm Laser surface scan

14 2013 Smith Global thresholding Mimics Hip (6) + shoulder (6) joint

Cadaver MDCT (GE Lightspeed+XCR/16)

0.63 mm / - Laser surface scan

69 2013 Yokota SSM Custom Femur (200) + acet. (200)

In-vivo MDCT Manual segmentation

45 2013 Zhou 3D adaptive thresholding MATLAB Hip joint (70) In-vivo MDCT (GE Toshiba) 1.5 mm / 0.73 mm

Manual segmentation

69 2012 Oneill Semi-automatic morphological snakes

Custom Femur (6) In-vivo MDCT (GE Lightspeed Plus) 1.25 mm / 0.68 mm

Manual segmentation

58 2011 Anstey Global thresholding Mimics Femur (1) + acet. (1)

Cadaver MDCT (GE LightSpeed+ XCR/16)

0.63 / - Laser surface scan

36 2011 Rathnayaka Global thresholding (A) / multi-thresholding (B) / edge detection (C)

MATLAB & Amira Femur (5) Cadaver MDCT (Philips Brilliance/64) 140 kVp / - 0.5 mm / 0.39 mm

3D contact scanner

54 2010 Thorhauer Manual Mimics Knee (2) Cadaver MDCT (GE Lightspeed/64) 1 mm / 0.80 mm

Laser surface scan

63 2009 Kainmueller SSM Custom Femur (30) + acet. (50)

In-vivo MDCT 0.5 mm - 5mm / 0.5 mm - 0.9 mm

Manual segmentation

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

73

5

Table 1. Publications included in this review.

Ref Year Author Segmentation method Software Anatomy (#) MaterialCT scanner (manufacturer/# slices)

CT tube voltage / tube current

CT slice thickness / voxel size Ground truth

55 2016 Szymor Global thresholding Slicer Skull (1) Dry bone CBCT (Accuitomo) 90 kV / 5 mA 0.5 mm / 0.25 mm

Optical 3D scan

61 2015 Chu Random forest & multi-atlas & articulated SSM

Custom & Amira Femur (60) + acet. (60)

In-vivo MDCT 1.6 mm / 0.6 mm

Manual segmentation

48 2015 Poleti Global thresholding Dolphin (A)/ InVesalius (B)

Mandible (10) Dry bone CBCT (i-CAT) 120 kV / 18.5 mAs - / 0.3 mm Digital calipers

64 2014 Balestra Articulated SSM Custom & Amira Femur (26) + pelvis (26)

In-vivo MDCT Manual segmentation

56 2014 Fasel Global thresholding n.a. (3 institutes) Skull (3) Cadaver MDCT (Siemens Sensation/64) Optical 3D scan49 2014 Kang Global thresholding InVivoDental (A)

/ Mimics (B) / OnDemand3D (C) / OsiriX (D)

Skull (3) Dry bone MDCT (Siemens Sensation/64) 0.6 mm / 0.44 mm

Optical 3D scan

46 2014 Santolaria Global thresholding (A) / Advanced thresholding (B)

Mimics Mandible (1) Dry bone CBCT (Kodak 9500) - / 0.2 mm - 0.3 mm

CMM

57 2014 Van den Broeck Global thresholding Mimics Tibia (10) Cadaver MDCT (GE LightSpeed VCT) 120 kVp / 160 mAs

0.63 mm / 0.39 mm

Optical 3D scan

44 2014 Zhang Region-based SSM Custom & Stradwin Femur (41) In-vivo MDCT 1.6 mm / 1.0 mm

Manual segmentation

60 2013 Chang Base invariant wavelet active shape model

Custom Maxilla (19) In-vivo CBCT (Sirona) - / 0.29 mm Manual segmentation

47 2013 Engelbrecht “Commercial company” (A) / “doctor” (B)

Mimics & SimPlant Ortho Pro

Mandible (7) Cadaver CBCT (KaVo 3D exam) - / 0.3 mm Laser surface scan

14 2013 Smith Global thresholding Mimics Hip (6) + shoulder (6) joint

Cadaver MDCT (GE Lightspeed+XCR/16)

0.63 mm / - Laser surface scan

69 2013 Yokota SSM Custom Femur (200) + acet. (200)

In-vivo MDCT Manual segmentation

45 2013 Zhou 3D adaptive thresholding MATLAB Hip joint (70) In-vivo MDCT (GE Toshiba) 1.5 mm / 0.73 mm

Manual segmentation

69 2012 Oneill Semi-automatic morphological snakes

Custom Femur (6) In-vivo MDCT (GE Lightspeed Plus) 1.25 mm / 0.68 mm

Manual segmentation

58 2011 Anstey Global thresholding Mimics Femur (1) + acet. (1)

Cadaver MDCT (GE LightSpeed+ XCR/16)

0.63 / - Laser surface scan

36 2011 Rathnayaka Global thresholding (A) / multi-thresholding (B) / edge detection (C)

MATLAB & Amira Femur (5) Cadaver MDCT (Philips Brilliance/64) 140 kVp / - 0.5 mm / 0.39 mm

3D contact scanner

54 2010 Thorhauer Manual Mimics Knee (2) Cadaver MDCT (GE Lightspeed/64) 1 mm / 0.80 mm

Laser surface scan

63 2009 Kainmueller SSM Custom Femur (30) + acet. (50)

In-vivo MDCT 0.5 mm - 5mm / 0.5 mm - 0.9 mm

Manual segmentation

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

74

5

Table 1. (continued)

Ref Year Author Segmentation method Software Anatomy (#) MaterialCT scanner (manufacturer/# slices)

CT tube voltage / tube current

CT slice thickness / voxel size Ground truth

70 2009 Ramme Expectation maximization BRAINS2 Phalanges (15) Cadaver MDCT (Siemens Sensation/64) 120 kVp / 105 mA

0.4 mm / 0.34 mm

Laser surface scan

42 2009 Zhang 3D adaptive thresholding Custom Calcaneus (1) + spine (2)

In-vivo MDCT (Toshiba HR / Siemens) 0.7 mm - 0.4 mm / 0.8 mm - 0.49 mm

Manual segmentation

53 2008 DeVries Manual BRAINS2 Phalanges (15) Cadaver MDCT (Siemens Sensation/64) 120 kVp / 105 mA 0.4 mm / 0.34 mm

Laser surface scan

41 2008 Gassman Artificial neural network BRAINS2 Phalanges (15) Cadaver MDCT (Siemens Sensation/64) 120 kVp / 105 mA 0.4 mm / 0.34 mm

Laser surface scan

59 2008 Gelaude Global thresholding (A) / self-developed contour-based algorithm (B)

Mimics Femur (15) Cadaver MDCT (Siemens Sensation) - / 1 mm Optical 3D scan

62 2008 Seim SSM Custom Pelvis (50) In-vivo MDCT 5 mm / 0.9 mm Manual segmentation

51 2006 Gelaude Adaptive contouring Mimics & MATLAB Pelvis (1) + humerus (1)

Cadaver MDCT 2 mm / - Manual segmentation

52 2006 Loubele Global thresholding Custom Skull (1) Skull phantom MDCT (Siemens Sensation/16) (A) / CBCT (I-CAT) (B)

Laser surface scan

66 2006 Rueda Active appearance model Custom Mandible (215) In-vivo MDCT Manual segmentation

60 2005 Schmidt Global thresholding / Self-developed border tracing algorithm *

Mimics / Custom Femur (1) + acet. (1)

Cryosectioned cadaver

MDCT 120 kV / 120 mA – 200 mA

1 mm / - Photographic images

67 2004 Lamecker SSM Custom Pelvis (23) In-vivo MDCT 5 mm / 1.4 mm Manual segmentation

43 2004 Rajamani SSM Custom Femur (14) In-vivo MDCT Manual segmentation

29 2002 Choi Global thresholding V-works Skull (1) Dry bone MDCT (Siemens Plus/4) 120 kVp / 200 mA 1 mm / - Anatomical landmarks

Acet. = acetabulum; CMM = coordinate measurement machine; MDCT = multi-detector row computed tomography; CBCT = cone-beam computed tomography. * Schmidt et al. (2005) did not disclose the identity of the segmentation methods used.

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

75

5

Table 1. (continued)

Ref Year Author Segmentation method Software Anatomy (#) MaterialCT scanner (manufacturer/# slices)

CT tube voltage / tube current

CT slice thickness / voxel size Ground truth

70 2009 Ramme Expectation maximization BRAINS2 Phalanges (15) Cadaver MDCT (Siemens Sensation/64) 120 kVp / 105 mA

0.4 mm / 0.34 mm

Laser surface scan

42 2009 Zhang 3D adaptive thresholding Custom Calcaneus (1) + spine (2)

In-vivo MDCT (Toshiba HR / Siemens) 0.7 mm - 0.4 mm / 0.8 mm - 0.49 mm

Manual segmentation

53 2008 DeVries Manual BRAINS2 Phalanges (15) Cadaver MDCT (Siemens Sensation/64) 120 kVp / 105 mA 0.4 mm / 0.34 mm

Laser surface scan

41 2008 Gassman Artificial neural network BRAINS2 Phalanges (15) Cadaver MDCT (Siemens Sensation/64) 120 kVp / 105 mA 0.4 mm / 0.34 mm

Laser surface scan

59 2008 Gelaude Global thresholding (A) / self-developed contour-based algorithm (B)

Mimics Femur (15) Cadaver MDCT (Siemens Sensation) - / 1 mm Optical 3D scan

62 2008 Seim SSM Custom Pelvis (50) In-vivo MDCT 5 mm / 0.9 mm Manual segmentation

51 2006 Gelaude Adaptive contouring Mimics & MATLAB Pelvis (1) + humerus (1)

Cadaver MDCT 2 mm / - Manual segmentation

52 2006 Loubele Global thresholding Custom Skull (1) Skull phantom MDCT (Siemens Sensation/16) (A) / CBCT (I-CAT) (B)

Laser surface scan

66 2006 Rueda Active appearance model Custom Mandible (215) In-vivo MDCT Manual segmentation

60 2005 Schmidt Global thresholding / Self-developed border tracing algorithm *

Mimics / Custom Femur (1) + acet. (1)

Cryosectioned cadaver

MDCT 120 kV / 120 mA – 200 mA

1 mm / - Photographic images

67 2004 Lamecker SSM Custom Pelvis (23) In-vivo MDCT 5 mm / 1.4 mm Manual segmentation

43 2004 Rajamani SSM Custom Femur (14) In-vivo MDCT Manual segmentation

29 2002 Choi Global thresholding V-works Skull (1) Dry bone MDCT (Siemens Plus/4) 120 kVp / 200 mA 1 mm / - Anatomical landmarks

Acet. = acetabulum; CMM = coordinate measurement machine; MDCT = multi-detector row computed tomography; CBCT = cone-beam computed tomography. * Schmidt et al. (2005) did not disclose the identity of the segmentation methods used.

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

76

5

Table 2. Mean and standard deviation of the reported accuracies of different CT image segmentation methods. Cad = cadaver, MSD = mean surface distance; RMSD = root-mean-square distance; MAD = mean absolute difference; MD = mean difference.

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

77

5

implants and complications during surgery. [23] Moreover, recent advances in computer-

assisted surgery, such as surgical navigation, [76] augmented reality, [77] and robot-guided

surgery, [78] will require highly accurate 3D surface models (< 1 mm). In this regard, it

should be noted that accuracy also depends on the resolution of the CT scanner. Currently,

routine diagnostic scan protocols typically have slice thicknesses ranging between 0.6

mm and 2 mm and voxel sizes ranging from 0.2 mm to 0.6 mm, which subsequently limits

the accuracy that can be achieved in clinical settings.

Suggestions for improvementThe limitations of the thresholding and SSM-based CT image segmentation methods

commonly used in medical AM are currently impeding the cost-effectiveness of patient-

specific AM constructs due to the overall time required for the segmentation and manual

post-processing. Therefore, more advanced methods are necessary that can accurately

segment different anatomical structures and their morphological variations. Therefore,

the authors of this study suggest that algorithms such as morphological snakes, [69]

artificial neural networks, [41] and expectation maximization [70] could markedly optimize

CT image segmentation in medical AM. In particular, deep convolutional neural networks

(CNNs) can be taught to obtain mid- and high-level abstractions from CT images. [79]

Based on the reviewed publications, the authors of this study suggest that future

research should focus on the clinical translation of advanced segmentation methods and

intelligent user interfaces. Such a novel interface would facilitate local segmentation using

multiple threshold values derived from local image histograms, edge detection algorithms,

and CNNs. Furthermore, an engineer should be able to interact with the segmentation

process in real time based on a preview of the 3D surface model. Such an approach will,

however, only be possible with advancements in graphical computing power.

Limitations of the reviewed studiesA major limitation of the reviewed publications was the incomplete specifications of

the CT data. A total of eleven publications (34%) did not mention the CT slice thickness

and/or the voxel size (see Table 1). Furthermore, in twenty-three publications (72%) the CT

tube voltage and tube current were not mentioned, and in eleven publications (34%)

the manufacturer of the CT scanner was not specified.

Another challenge faced when evaluating the publications was the diversity of metrics

used to assess the accuracies. A sizable body of work reported accuracies using relative

distance- based metrics such as volume percentages. [80] These publications were

not included in this review. Only publications that used mean distance-based accuracy

metrics in millimeters (mean surface distance, root-mean-square distance, mean absolute

difference and mean difference) were included. However, it should be noted that the root-

mean-square distance is frequently misinterpreted in practice, and that the mean absolute

difference and mean difference are commonly only based on a limited number of linear

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

78

5

measurements. We therefore recommend that the geometric accuracy of bone 3D surface

models be evaluated using the mean surface distance. Nevertheless, when only the mean

surface distance is reported, large deviations in local anatomical regions can either remain

unnoticed or can strongly influence the mean surface distance. [67] Therefore, the authors

of this study recommend that an additional surface comparison is provided using colors to

depict anatomy-specific deviations (a “heat map”). Such heat maps were used in eleven

publications (34%).

Another limitation of the reviewed studies was the lack of information on the amount

of manual editing that was required in order to achieve the reported accuracies. Manual

editing inherently introduces intra- and inter-operator variability. Therefore, we suggest

that the experience of the operator, the exact nature of manual editing, and the time spent

should always be specified in detail.

Finally, numerous studies included in this review acquired high-resolution CT scans

using dry bones (5 publications) or cadaver specimens (12 publications), which resulted

in mean accuracies of 0.30 mm ± 0.30 mm and 0.46 mm ± 0.35 mm, respectively. In-vivo

CT images (13 publications) resulted in a mean accuracy of 0.73 mm ± 0.36 mm. One

explanation for these differences is that the photon energy spectrum of an X-ray beam

changes as it passes through soft tissues, [81] which results in differences in absorption and

scattering during CT imaging. Moreover, the gray values of soft tissues are closer to those

of bone than air. Consequently, soft tissues make semi-automated bone segmentation

less accurate. [82] Therefore, the reported accuracies found in those studies using dry

bones are not translatable to clinical conditions. Furthermore, musculoskeletal bones

in humans differ markedly in morphology, which causes differences in the accuracy and

manual clean-up time. For example, segmentation of small scaffold-like orbital bones is

more challenging than that of large femoral bones.

Taking the above into account, we hypothesize that the best method to evaluate

the accuracy of medical AM constructs is to use multiple life-like cadaveric specimens

and to subsequently remove all soft tissues from the bone after CT imaging. This offers

the possibility to acquire an optical 3D surface scan of the bone that can be used as an

accurate ground truth (<0.05 mm). Note that boiling or chemical treatments should be

avoided since these treatments can introduce geometric deviations to the bone of up to

0.49 mm [59] and 0.43 mm, [57] respectively.

CONCLUSIONThis literature review revealed that a plethora of different CT image segmentation methods

are currently used for bone segmentation. The accuracy of these image segmentation

methods differed markedly. Global thresholding remains the most widely used CT

image segmentation method in medical AM but often requires extensive manual post-

processing. Advanced thresholding approaches could improve the accuracy of global

thresholding, but such methods are currently not implemented in commercially available

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

79

5

software packages. The development of fully automatic and adaptive image segmentation

algorithms could improve the accuracy and reduce the costs of patient-specific AM

constructs. More specifically, future research should focus on the development of image

segmentation methods using CNNs.

COMPETING INTERESTSNone declared

FUNDINGNone

ETHICAL APPROVALNot required

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

80

5

REFERENCES1. Wurm G, Lehner M, Tomancok B,

Kleiser R, and Nussbaumer K (2011) Cerebrovascular Biomodeling for Aneurysm Surgery: Simulation-Based Training by Means of Rapid Prototyping Technologies. Surg Innov 18:294–306

2. Mitsouras D, Liacouras P, Imanzadeh A, Giannopoulos A A, Cai T, Kumamaru K K, George E, Wake N, Caterson E J, Pomahac B, Ho V B, Grant G T, and Rybicki F J (2015) Medical 3D Printing for the Radiologist. Radiographics 35:1965–1988

3. D’Urso P S, Barker T M, Earwaker W J, Bruce L J, Atkinson R L, Lanigan M W, Arvier J F, and Effeney D J (1999) Stereolithographic biomodelling in cranio-maxillofacial surgery: a prospective trial. J. Cranio-Maxillofacial Surg. 27:30–37

4. Tam M D, Laycock S D, Bell D G, and Chojnowski A (2012) 3-D printout of a DICOM file to aid surgical planning in a 6 year old patient with a large scapular osteochondroma complicating congenital diaphyseal aclasia. J. Radiol. Case Rep. 6:31–37

5. McMenamin P G, Quayle M R, McHenry C R, and Adams J W (2014) The production of anatomical teaching resources using three-dimensional (3D) printing technology. Anat. Sci. Educ. 7:479–486

6. Bakhos D, Velut S, Robier A, Al Zahrani M, and Lescanne E (2010) Three-Dimensional Modeling of the Temporal Bone for Surgical Training. Otol. Neurotol. 31:328–334

7. Dawood A, Marti B M, Sauret-Jackson V, and Darwood A (2015) 3D printing in dentistry. Br Dent J 219:521–529

8. Ramasamy M, Giri, Raja R, Subramonian, Karthik, and Narendrakumar R (2013) Implant surgical guides: From the past to the present. J. Pharm. Bioallied Sci. 5:S98–S102

9. Jardini A L, Larosa M A, Maciel Filho R, Zavaglia C A D C, Bernardes L F, Lambert C S, Calderoni D R, and Kharmandayan P (2014) Cranial reconstruction: 3D biomodel and custom-built implant created using additive manufacturing. J. Cranio-Maxillofacial Surg. 42:1877–1884

10. Ibrahim D, Broilo T L, Heitz C, de Oliveira M G, de Oliveira H W, Nobre S M W, dos Santos Filho J H G, and Silva D N (2009) Dimensional error of selective laser sintering, three-dimensional printing and PolyJetTM models in the reproduction of mandibular anatomy. J. Cranio-Maxillofacial Surg. 37:167–173

11. Martelli N, Serrano C, van den Brink H, Pineau J, Prognon P, Borget I, and El Batti S (2016) Advantages and disadvantages of 3-dimensional printing in surgery: A systematic review. Surgery 159:1485–1500

12. Bibb R, Eggbeer D, and Paterson A (2015) Medical modelling: The application of advanced design and rapid prototyping techniques in medicine. 2nd ed. Woodhead Publishing

13. Ibrahim D, Broilo T L, Heitz C, de Oliveira M G, de Oliveira H W, Nobre S M W, Dos Santos Filho J H G, and Silva D N (2009) Dimensional error of selective laser sintering, three-dimensional printing and PolyJet models in the reproduction of mandibular anatomy. J. Craniomaxillofac. Surg. 37:167–73

14. Smith E J, Anstey J A, Venne G, and Ellis R E (2013) Using additive manufacturing in accuracy evaluation of reconstructions from computed tomography. Proc Inst Mech Eng H 227:551–559

15. Huotilainen E, Paloheimo M, Salmi M, Paloheimo K-S, Björkstrand R, Tuomi J, Markkola A, and Mäkitie A (2014) Imaging requirements for medical applications of additive manufacturing. Acta radiol. 55:78–85

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

81

5

16. Liang X, Lambrichts I, Sun Y, Denis K, Hassan B, Li L, Pauwels R, and Jacobs R (2010) A comparative evaluation of Cone Beam Computed Tomography (CBCT) and Multi-Slice CT (MSCT). Part II: On 3D model accuracy. Eur. J. Radiol. 75:270–274

17. Koivisto J, Kiljunen T, Kadesjö N, Shi X-Q, and Wolff J (2015) Effective radiation dose of a MSCT, two CBCT and one conventional radiography device in the ankle region. J. Foot Ankle Res. 8

18. Pauwels R, Beinsberger J, Collaert B, Theodorakou C, Rogers J, Walker A, Cockmartin L, Bosmans H, Jacobs R, Bogaerts R, and Horner K (2012) Effective dose range for dental cone beam computed tomography scanners. Eur. J. Radiol. 81:267–271

19. Winder J and Bibb R (2005) Medical Rapid Prototyping Technologies: State of the Art and Current Limitations for Application in Oral and Maxillofacial Surgery. J. Oral Maxillofac. Surg. 63:1006–1015

20. Hughes CW, Page K, Bibb R, Taylor J, Revington P. The custom-made titanium orbital floor prosthesis in reconstruction for orbital floor fractures. British Journal of Oral and Maxillofacial Surgery 2003;41:50-53.

21. Hiller J D and Lipson H (2009) STL 2.0: a proposal for a universal multi-material Additive Manufacturing File format. In: Proceedings of the Solid Freeform Fabrication Symposium, 266–278

22. Huotilainen E, Jaanimets R, Valášek J, Marcián P, Salmi M, Tuomi J, Mäkitie A, and Wolff J (2014) Inaccuracies in additive manufactured medical skull models caused by the DICOM to STL conversion process. J. Cranio-Maxillofacial Surg. 42:e259–e265

23. Stoor P, Suomalainen A, Lindqvist C, Mesimaki K, Danielsson D, Westermark A, and Kontio R K (2014) Rapid prototyped patient specific implants for reconstruction of orbital wall defects. J. cranio-maxillo-facial Surg. 42:1644–1649

24. Fourie Z, Damstra J, Schepers R H, Gerrits P O, and Ren Y (2012) Segmentation process significantly influences the accuracy of 3D surface models derived from cone beam computed tomography. Eur. J. Radiol. 81:e524–e530

25. Sharma N and Aggarwal L M (2010) Automated medical image segmentation techniques. J. Med. Phys. 35:3–14

26. Hieu L C, Zlatov N, Sloten J Vander, Bohez E, Khanh L, Binh P H, Oris P, and Toshev Y (2005) Medical rapid prototyping applications and methods. Assem. Autom. 25:284–292

27. Rengier F, Mehndiratta A, von Tengg-Kobligk H, Zechmann C M, Unterhinninghofen R, Kauczor H-U, and Giesel F L (2010) 3D printing based on imaging data: review of medical applications. Int. J. Comput. Assist. Radiol. Surg. 5:335–341

28. Gibson I, Cheung L K, Chow S P, Cheung W L, Beh S L, Savalani M, and Lee S H (2006) The use of rapid prototyping to assist medical applications. Rapid Prototyp. J. 12:53–58

29. Choi J Y, Choi J H, Kim N K, Kim Y, Lee J K, Kim M K, Lee J H, and Kim M J (2002) Analysis of errors in medical rapid prototyping models. Int. J. Oral Maxillofac. Surg. 31:23–32

30. Klein H M, Schneider W, Alzen G, Voy E D, and Günther R W (1992) Pediatric craniofacial surgery: Comparison of milling and stereolithography for 3D model manufacturing. Pediatr. Radiol. 22:458–460

31. Otsu N (1975) A Threshold Selection Method from Gray-Level Histograms. Automatica 20:62–66

32. Canny J (1986) A Computational Approach to Edge Detection. IEEE Trans. Pattern Anal. Mach. Intell. PAMI-8:679–698

33. Adams R and Bischof L (1994) Seeded region growing. IEEE Trans. Pattern Anal. Mach. Intell. 16:641–647

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

82

5

34. van Eijnatten M, Koivisto J, Karhu K, Forouzanfar T, and Wolff J (2016) The impact of manual threshold selection in medical additive manufacturing. Int. J. Comput. Assist. Radiol. Surg. 12:607–615

35. Rathnayaka K, Momot K I, Noser H, Volp A, Schuetz M A, Sahama T, and Schmutz B (2014) Quantification of the accuracy of MRI generated 3D models of long bones compared to CT generated 3D models. Med. Eng. Phys. 34:357–363

36. Rathnayaka K, Sahama T, Schuetz M A, and Schmutz B (2011) Effects of CT image segmentation methods on the accuracy of long bone 3D reconstructions. Med. Eng. Phys. 33:226–33

37. Zhang J, Yan C H, Chui C K, and Ong S H (2010) Fast segmentation of bone in CT images using 3D adaptive thresholding. Comput. Biol. Med. 40:231–236

38. Withey D J and Koles Z J (2007) Medical Image Segmentation: Methods and Software. 2007 Jt. Meet. 6th Int. Symp. Noninvasive Funct. Source Imaging Brain Hear. Int. Conf. Funct. Biomed. Imaging (NFSI-ICFBI 2007): 140–143

39. Heimann T and Meinzer H P (2009) Statistical shape models for 3D medical image segmentation: A review. Med. Image Anal. 13:543–563

40. Aspert N, Santa-Cruz D, and Ebrahimi T (2002) MESH: measuring errors between surfaces using the Hausdorff distance. In: Proceedings. IEEE International Conference on Multimedia and Expo. 705–708 vol.1

41. Gassman E E, Powell S M, Kallemeyn N A, Devries N A, Shivanna K H, Magnotta V A, Ramme A J, Adams B D, and Grosland N M (2008) Automated bony region identification using artificial neural networks: reliability and validation measurements. Skeletal Radiol. 37:313–319

42. Zhang J (2009) Model-Based Segmentation and Registration of Multimodal Medical Images, National University of Singapore

43. Rajamani K, Nolte L, and Styner M (2004) A Novel Approach to Anatomical Structure Morphing for Intraoperative Visualization. In: Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2004: 7th International Conference, Saint-Malo, France, September 26-29, 2004. Proceedings, Part II C. Barillot, D. R. Haynor, and P. Hellier, Eds. Berlin, Heidelberg, 478–485

44. Zhang J, Malcolm D, Hislop-Jambrich J, Thomas C D L, and Nielsen P M F (2014) An anatomical region-based statistical shape model of the human femur. Comput. Methods Biomech. Biomed. Eng. Imaging Vis. 2:176–185

45. Zhou S, Cheng Y, Wang Y, Dong K, Guo C, Bai J, and Tamura S (2013) Segmentation of the hip joint in CT volumes using adaptive thresholding classification and normal direction correction. J. Chinese Inst. Eng. 36:1059–1072

46. Santolaria J, Jiménez R, Rada M, and Loscos F (2014) Error compensation method for improving the accuracy of biomodels obtained from CBCT data. Med. Eng. Phys. 36:397–404

47. Engelbrecht W P, Fourie Z, Damstra J, Gerrits P O, and Ren Y (2013) The influence of the segmentation process on 3D measurements from cone beam computed tomography-derived surface models. Clin. Oral Investig. 17:1919–1927

48. Poleti M L, Fernandes T M F, Pagin O, Moretti M R, and Rubira-Bullen I R F (2016) Analysis of linear measurements on 3D surface models using CBCT data segmentation obtained by automatic standard pre-set thresholds in two segmentation software programs: an in vitro study. Clin. Oral Investig. 20:179–185

49. Kang S-H, Kim M-K, Kim H-J, Zhengguo P, and Lee S-H (2014) Accuracy Assessment of Image-Based Surface Meshing for Volumetric Computed Tomography

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

83

5

Images in the Craniofacial Region. J. Craniofac. Surg. 25:2051–2055

50. Schmidt J, Engh J, Viceconti M, and Ploeg H (2005) What is the accuracy of surface models created from visible human male computed tomography data? ASB 29th Annu. Meet., p. 63

51. Gelaude F, Sloten J Vander, and Lauwers B (2006) Semi-automated segmentation and visualisation of outer bone cortex from medical images. Comput. Methods Biomech. Biomed. Engin. 9:65–77

52. Loubele M, Maes F, Vandermeulen D, Denis K, Jacobs R, White S C, van Steenberghe D, Van Bael A, Loeckx D, Lambrichts I, and Suetens P (2006) Assessment of bone segmentation quality of CT scanners using laser scanning. Int J CARS 1:389–402

53. DeVries N a, Gassman E E, Kallemeyn N a, Shivanna K H, Magnotta V a, and Grosland N M (2008) Validation of phalanx bone three-dimensional surface segmentation from computed tomography images using laser scanning. Skeletal Radiol. 37:35–42

54. Thorhauer E, Miyawaki M, Illingworth K, Holmes J A, and Anderst W (2010) Accuracy of bone and cartilage models obtained from CT and MRI. In: Poster presented at the 34th Annual Meeting of the American Society of Biomechanics. 18–21

55. Szymor P, Kozakiewicz M, and Olszewski R (2016) Accuracy of open-source software segmentation and paper-based printed three-dimensional models. J. Cranio-Maxillofacial Surg. 44:202–209

56. Fasel J H D, Beinemann J, Schaller K, and Peitgen H-O (2014) Computer Science Tools for Manual Editing of Computed Tomographic Images: Impact on the Quality of 3D Printed Models. Surg. Sci. 5:439–443

57. Van den Broeck J, Vereecke E, Wirix-Speetjens R, and Vander Sloten J (2014)

Segmentation accuracy of long bones. Med. Eng. Phys. 36:949–953

58. Anstey J B, Smith E J, Rasquinha B, Rudan J F, and Ellis R E (2011) On the use of laser scans to validate reverse engineering of bony anatomy. Stud. Health Technol. Inform. 163:18–24

59. Gelaude F, Vander Sloten J, and Lauwers B (2008) Accuracy assessment of CT-based outer surface femur meshes. Comput. Aided Surg. 13:188–199

60. Chang Y-B, Xia J J, Yuan P, Kuo T-H, Xiong Z, Gateno J, and Zhou X (2013) 3D segmentation of maxilla in cone-beam computed tomography imaging using base invariant wavelet active shape model on customized two-manifold topology. J. Xray. Sci. Technol. 21:251–282

61. Chu C, Chen C, Liu L, and Zheng G (2015) FACTS: Fully Automatic CT Segmentation of a Hip Joint. Ann. Biomed. Eng. 43:1247–1259

62. Seim H, Kainmueller D, Heller M, Lamecker H, Zachow S, and Hege H-C (2008) Automatic Segmentation of the Pelvic Bones from CT Data Based on a Statistical Shape Model. In: Proceedings of the First Eurographics Conference on Visual Computing for Biomedicine, 93–100

63. Kainmueller D, Lamecker H, Zachow S, and Hege H-C (2009) An articulated statistical shape model for accurate hip joint segmentation. Conf. Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. 6345–6351

64. Balestra S, Schumann S, Heverhagen J, Nolte L, and Zheng G (2014) Articulated Statistical Shape Model-Based 2D-3D Reconstruction of a Hip Joint. In: Information Processing in Computer-Assisted Interventions: 5th International Conference, IPCAI 2014, Fukuoka, Japan, June 28, 2014. Proceedings D. Stoyanov, D. L. Collins, I. Sakuma, P. Abolmaesumi, and P. Jannin, Eds. Cham, 128–137

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

84

5

65. Rajamani K T, Joshi S C, and Styner M A (2004) Bone model morphing for enhanced surgical visualization. In Biomedical Imaging: Nano to Macro, 2004. IEEE International Symposium on. 1255–1258

66. Rueda S, Gil J A, Pichery R, and Alcañiz M (2006) Automatic Segmentation of Jaw Tissues in CT Using Active Appearance Models and Semi-automatic Landmarking. In: Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2006: 9th International Conference, Copenhagen, Denmark, October 1-6, 2006. Proceedings, Part I R. Larsen, M. Nielsen, and J. Sporring, Eds. Berlin, Heidelberg, 167–174

67. Lamecker H, Seebass M, Hege H-C, and Deuflhard P (2004) A 3D statistical shape model of the pelvic bone for segmentation. Proc. SPIE 5370:1341–1351

68. Yokota F, Okada T, Takao M, Sugano N, Tada Y, Tomiyama N, and Sato Y (2013) Automated CT segmentation of diseased hip using hierarchical and conditional statistical shape models. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. 190–197

69. O’Neill G T, Lee W-S, and Beaulé P (2011) Segmentation of cam-type femurs from CT scans. Vis. Comput. 28:205–218

70. Ramme A J, DeVries N, Kallemyn N a, Magnotta V a, and Grosland N M (2009) Semi-automated phalanx bone segmentation using the expectation maximization algorithm. J. Digit. Imaging 22:483–91

71. Pinto J M, Arrieta C, Andia M E, Uribe S, Ramos-Grez J, Vargas A, Irarrazaval P, and Tejos C (2015) Sensitivity analysis of geometric errors in additive manufacturing medical models. Med. Eng. Phys. 37:328–334

72. EU (2017) Revisions of medical device directives - European Commission. Available: https://ec.europa.eu/growth/sectors/medical-devices/regulatory-

framework/revision_en. [Accessed: 30-May-2017]

73. Zhu Z and Li G (2011) Construction of 3D Human Distal Femoral Surface Models Using a 3D Statistical Deformable Model. J. Biomech. 44:2362–2368

74. Vos W De, Casselman J, and Swennen G R J (2009) Cone-beam computerized tomography (CBCT) imaging of the oral and maxillofacial region: A systematic review of the literature. Int. J. Oral Maxillofac. Surg. 38:609–625

75. Ventola C L (2014) Medical Applications for 3D Printing : Current and Projected Uses. Pharm. Ther. 39:704–711

76. He Y, Zhang Y, An J, Gong X, Feng Z, and Guo C (2013) Zygomatic surface marker-assisted surgical navigation: a new computer-assisted navigation method for accurate treatment of delayed zygomatic fractures. J. Oral Maxillofac. Surg. 71:2101–2114

77. Badiali G, Ferrari V, Cutolo F, Freschi C, Caramella D, Bianchi A, and Marchetti C (2014) Augmented reality as an aid in maxillofacial surgery: validation of a wearable system allowing maxillary repositioning. J. cranio-Maxillofacial Surg. 42:1970–1976

78. Majdani O, Rau T S, Baron S, Eilers H, Baier C, Heimann B, Ortmaier T, Bartling S, Lenarz T, and Leinung M (2009) A robot-guided minimally invasive approach for cochlear implant surgery: preliminary results of a temporal bone study. Int. J. Comput. Assist. Radiol. Surg. 4:475–486

79. Greenspan H, Ginneken B van, and Summers R M (2016) Guest Editorial Deep Learning in Medical Imaging: Overview and Future Promise of an Exciting New Technique. IEEE Trans. Med. Imaging 35:1153–1159

80. Udupa J K, Leblanc V R, Zhuge Y, Imielinska C, Schmidt H, Currie L M, Hirsch B E, and Woodburn J (2006) A framework for evaluating image segmentation algorithms. Comput. Med. Imaging Graph. 30:75–87

CT IM

AG

E SE

GM

EN

TATIO

N M

ETH

OD

S FOR

BO

NE

USE

D IN

ME

DIC

AL A

DD

ITIVE

MA

NU

FAC

TUR

ING

85

5

81. Barrett J F and Keat N (2004) Artifacts in CT: Recognition and Avoidance. RadioGraphics 24:1679–1691

82. van Eijnatten M, Berger F H, de Graaf P, Koivisto J, Forouzanfar T, and Wolff J (2017) Influence of CT parameters on STL model accuracy. Rapid Prototyp. J. 23:679–685

6 THE IMPACT OF MANUAL THRESHOLD SELECTION IN

MEDICAL ADDITIVE MANUFACTURING

Maureen van Eijnatten, Juha Koivisto, Kalle Karhu,

Tymour Forouzanfar, Jan Wolff

International Journal of Computer Assisted Radiology and Surgery (2016). 12(4): 607-615doi: 10.1007/s11548-016-1490-4

THE

IMPA

CT O

F MA

NU

AL TH

RE

SHO

LD SE

LEC

TION

IN M

ED

ICA

L AD

DITIV

E M

AN

UFA

CTU

RIN

G

88

6

ABSTRACTPurposeMedical additive manufacturing requires standard tessellation language (STL) models.

Such models are commonly derived from computed tomography (CT) images using

thresholding. Threshold selection can be performed manually or automatically. The aim

of this study was to assess the impact of manual and default threshold selection on

the reliability and accuracy of skull STL models using different CT technologies.

MethodOne female and one male human cadaver head were imaged using multi-detector row

CT, dual-energy CT, and two cone-beam CT scanners. Four medical engineers manually

thresholded the bony structures on all CT images. The lowest and highest selected mean

threshold values and the default threshold value were used to generate skull STL models.

Geometric variations between all manually thresholded STL models were calculated.

Furthermore, in order to calculate the accuracy of the manually and default thresholded

STL models, all STL models were superimposed on an optical scan of the dry female and

male skulls (“gold standard”).

ResultsThe intra- and inter-observer variability of the manual threshold selection was good (intra-

class correlation coefficients >0.9). All engineers selected grey values closer to soft tissue

to compensate for bone voids. Geometric variations between the manually thresholded

STL models were 0.13 mm (multi-detector row CT), 0.59 mm (dual-energy CT), and 0.55

mm (cone-beam CT). All STL models demonstrated inaccuracies ranging from -0.8 mm

to +1.1 mm (multi-detector row CT), –0.7 mm to +2.0 mm (dual-energy CT), and -2.3 mm

to +4.8 mm (cone-beam CT).

ConclusionsThis study demonstrates that manual threshold selection results in better STL models

than default thresholding. The use of dual-energy CT and cone-beam CT technology in

its present form does not deliver reliable or accurate STL models for medical additive

manufacturing. New approaches are required that are based on pattern recognition and

machine learning algorithms.

THE

IMPA

CT O

F MA

NU

AL TH

RE

SHO

LD SE

LEC

TION

IN M

ED

ICA

L AD

DITIV

E M

AN

UFA

CTU

RIN

G

89

6

Figure 1. A schematic diagram of the three steps required to fabricate an AM medical construct.

INTRODUCTIONAdditive manufacturing (AM), also known as three-dimensional (3D) printing, refers to

a process where a series of successive layers are laid down to create a 3D construct. AM

combined with advanced medical imaging technologies such as computed tomography

(CT) and magnetic resonance imaging (MRI) has resulted in a paradigm shift in medicine

from traditional serial production to patient-specific constructs. This combination of

technologies offers new possibilities for the fabrication of implants, saw guides and drill

guides that are designed to meet the specific anatomical needs of patients [1].

The three-step medical AM process begins with image acquisition (Figure 1, Step 1),

which is commonly performed using a multi-detector row computed tomography (MDCT)

scanner. However, dual-energy computed tomography (DECT), which offers the possibility

of acquiring CT images using two different X-ray spectra, is becoming more common

in hospital environments [2]. Furthermore, cone-beam computed tomography (CBCT) is

being increasingly used in dentistry and maxillofacial surgery due to its low costs and

reduced radiation dose when compared with MDCT scanners [3].

Images acquired using CT technologies are commonly saved as Digital Imaging and

Communications in Medicine (DICOM) files. These files contain voxels with grey values that

are proportional to the attenuation coefficient in the corresponding volume of the patient.

In MDCT, these grey values are scaled according to Hounsfield units (HU): air (-1000 HU),

water (0 HU), and compact bone (+1000 HU). In CBCT technology, the degree of x-ray

attenuation is scaled using grey values, hence voxel values [4]. CBCT grey values are often

arbitrary and do not correspond to MDCT HU values [3], [5], [6]. Furthermore, a large

variability in the grey values has been reported between different CBCT scanners [7], [8].

At present, medical AM requires the conversion of DICOM images into virtual 3D

surface models that are commonly saved as standard tessellation language (STL) files

(Figure 1, Step 2). STL models are commonly used to design medical constructs using

computer-aided design (CAD) software. The DICOM-to-STL conversion process requires

the partitioning and hence the segmentation of voxels into different tissue types. The most

THE

IMPA

CT O

F MA

NU

AL TH

RE

SHO

LD SE

LEC

TION

IN M

ED

ICA

L AD

DITIV

E M

AN

UFA

CTU

RIN

G

90

6

common segmentation method used to date is thresholding. During the thresholding

process, all voxels with a grey value that is equal or greater than a selected threshold value

t are included in a segmented volume [9] using a binary mask Mx,y (Equation 1):

(1)

where Ix,y denotes the grey value at coordinates x and y.

The medical image segmentation software packages available offer only a single,

default threshold value for compact bone, soft tissue, and cartilage. However, these

default values are often not optimized for all types of MDCT, DECT, and CBCT images

and do not take into account the variations in grey values between different scanners [10].

Therefore, in most cases, manual threshold selection is necessary to acquire an optimal

STL model. Threshold selection, however, still remains a subjective task [11], especially in

the head area due to the plethora of complex bony geometries (Figure 2). Furthermore,

minor dislocations in the facial area can have an impact on patient function and

aesthetic appearance.

At present, there is a paucity of literature on threshold selection in the head area for

medical purposes. Therefore, the aim of this study was to assess the impact of manual and

default threshold selection on the reliability and accuracy of skull STL models acquired

using different MDCT and CBCT technologies.

MATERIALS AND METHODSOne female and one male human cadaver head were anonymously provided by

the Department of Anatomy, VU University Medical Center Amsterdam, The Netherlands.

The two heads were embedded in a novel embalming liquid “Fix for Life” [12] that produces

near life-like cadavers. Ethical approval for this study was provided by the Medical Ethical

Committee of the VU University Medical Center (Ref. 2016.401).

M��� � � 0I��� � �1I��� � � , ( 1 )

Figure 2. The effect of threshold selection on skull STL models.

THE

IMPA

CT O

F MA

NU

AL TH

RE

SHO

LD SE

LEC

TION

IN M

ED

ICA

L AD

DITIV

E M

AN

UFA

CTU

RIN

G

91

6

The two “Fix for Life” cadaver heads were imaged using the following CT technologies:

GE Discovery CT750 HD 64-slice MDCT (GE Healthcare, Little Chalfont, Buckinghamshire,

UK), NewTom 5G CBCT (NewTom, Verona, Italy), and Vatech PaX Zenith 3D CBCT (Vatech,

Fort Lee, USA) (Figure 3, Step 1). The GE Discovery CT750 MDCT scanner was also

operated using a dual-energy imaging mode (DECT). All scanners and image acquisition

parameters are summarized in Table 1.

After CT image acquisition, all DICOM files were imported into Osirix® MD software

(Osirix Foundation, Geneva, Switzerland). This software is FDA-cleared, CE-labeled for

primary diagnostics, and is commonly used in medical AM. Osirix® MD software provides

options for both manual and default threshold selection.

Four medical engineers were subsequently requested to manually select the optimal

threshold value for bone in order to create an accurate STL model of the female and

male skull, hence facial bony structures (Figure 3, Step 2). All four engineers were blinded

for their own results and those of others. The manual threshold selection procedure was

repeated after a six-week interval in order to determine the intra-observer variability and

to calculate the mean threshold value. In addition, the inter-observer variability and intra-

class correlation coefficients (ICC) were calculated using SPSS® software (SPSS® version 22,

Chicago, Il, USA). ICC ranges between 0 and 1, with 0 corresponding to no agreement and

1 corresponding to complete agreement [13].

In order to graphically represent the distribution of grey values in the manually selected

and default threshold values, histograms were plotted for each of the four CT scanners

Figure 3. Outline of the study.

THE

IMPA

CT O

F MA

NU

AL TH

RE

SHO

LD SE

LEC

TION

IN M

ED

ICA

L AD

DITIV

E M

AN

UFA

CTU

RIN

G

92

6

Fig

ure

3. O

utlin

e o

f the

stu

dy.

GE

Dis

cove

ry C

T750

HD

64

-slic

e (M

DC

T)G

E D

isco

very

CT7

50 H

D

64-s

lice

(DE

CT)

New

Tom

5G

(CB

CT)

Vate

ch P

aX Z

enit

h 3D

(C

BC

T)

fem

ale

mal

efe

mal

eM

ale

fem

ale

mal

efe

mal

em

ale

Tub

e vo

ltag

e (k

V)

120

120

80,1

4080

,140

110

110

115

115

Tub

e cu

rren

t (m

A)

300

300

375

375

7 7

66

Exp

osu

re t

ime

(s)

0.63

8 0.

912

0.91

20.

699

1010

2424

Spac

ing

bet

wee

n sl

ices

(mm

)0.

312

0.31

20.

312

0.31

2n.

a.n.

a.n.

a.n.

a.Sl

ices

thi

ckne

ss (m

m)

0.62

50.

625

0.62

50.

625

0.30

00.

300

0.30

00.

300

Num

ber

of v

o x

els

512

x 51

2 x

767

512

x 51

2 x

919

512

x 51

2 x

767

512

x 51

2 x

919

610

x 61

0 x

539

610

x 61

0 x

541

800

x 80

0 x

632

800

x 80

0 x

632

Rec

ons

truc

tion

kern

elB

one

plu

sB

one

plu

sSo

ftSo

ftSt

and

ard

Stan

dar

dn.

a.n.

a.

THE

IMPA

CT O

F MA

NU

AL TH

RE

SHO

LD SE

LEC

TION

IN M

ED

ICA

L AD

DITIV

E M

AN

UFA

CTU

RIN

G

93

6

using MatLab® software (MatLab v.2012, MathWorks, Natick, Massachusetts, USA) (Figure

4). Only the highest and lowest mean selected threshold values presented on the eight

histograms were used to generate STL models (Figure 3, Step 3). The generated STL

models were subsequently geometrically compared to each other using GOM Inspect®

software (GOM Inspect v8, GOM  mbH, Braunschweig, Germany) in order to calculate

the variations between the highest and lowest threshold STL models (Figure 3, Step 4).

In a final step, all soft tissues were manually removed from the cadaver heads using

standard dissection equipment (i.e., scrapers and scalpels) by a highly experienced

technician of the Department of Anatomy. Manual removal was opted for since this

procedure ensured minimal dimensional changes in the bony structures of the cadaver

skulls [14]. The resulting dry female and male skulls were subsequently scanned using

a GOM ATOS™ III optical 3D scanner (GOM  GmbH, Braunschweig, Germany) with an

accuracy of < 0.05 mm to acquire a “gold standard” STL model of the skulls (Figure 3).

These “gold standard” STL models were subsequently superimposed on the STL models

generated using the highest and lowest manually selected and default threshold values in

order to calculate the accuracy of each thresholded STL model (Figure 3, Step 5).

RESULTSThe intra- and inter-observer reliability results of all manually selected threshold values

are presented in Table 2. All selected threshold values ranged from 113 HU to 303 HU for

the MDCT and DECT technologies and from 537 gv to 1281 gv for the CBCT technologies

(Figure 4 A - H). As shown in the histograms, all the selected threshold values differed

from the default threshold value provided by Osirix MD® software (500 HU). Furthermore,

the geometric variations between the highest and lowest thresholded STL models were

larger in the STL models derived from DECT and CBCT when compared with the MDCT-

derived STL models (Figure 5).

When compared to the “gold standard”, all manually and automatically thresholded

STL models demonstrated inaccuracies ranging from -0.8 mm to +1.1 mm, –0.7 mm to

+2.0 mm, and -2.3 mm to +4.8 mm for all STL models derived from MDCT, DECT, and

CBCT, respectively (Figure 6 A - K). The male skull presented comparable accuracies to

those observed on the female skull. The MDCT and DECT-derived STL models acquired

using the default threshold value demonstrated the highest loss of bone-HU values (Figure

6 C, F). The NewTom CBCT-derived STL model acquired using the default threshold value

(500 HU) provided by Osirix MD software resulted in an increase in artifacts and noise

(Figure 6 I). The Vatech CBCT DICOM images did not allow the creation of an STL model

using the 500-HU default threshold value since the grey values were not scaled to HU

values (Figure 4 D,H).

THE

IMPA

CT O

F MA

NU

AL TH

RE

SHO

LD SE

LEC

TION

IN M

ED

ICA

L AD

DITIV

E M

AN

UFA

CTU

RIN

G

94

6

Table 2. Intra- and inter-observer variability of manual threshold selection by four medical engineers on CT images of a female and a male cadaver head.

Intra-observer variability Inter-observer variability between the engineers

Intra-class correlation coefficient (ICC)

ICCEngineer 2female / male

ICCEngineer 3female / male

ICCEngineer 4female / maleCadaver head female / male

Engineer 1 0.999 / 0.997 Engineer 1 0.994 / 0.988 0.980 / 0.987 0.970 / 0.954Engineer 2 0.995 / 0.995 Engineer 2 0.978 / 0.998 0.961 / 0.931Engineer 3 0.992 / 0.999 Engineer 3 0.914 / 0.917Engineer 4 0.969 / 0.989 Engineer 4

Figure 4 (A to H). The mean threshold values (HU) selected by four medical engineers and the pre-defined default threshold value (500 HU) are presented in histograms A to H. The y-axis of the histograms (frequencies) are set to a logarithmic scale.

THE

IMPA

CT O

F MA

NU

AL TH

RE

SHO

LD SE

LEC

TION

IN M

ED

ICA

L AD

DITIV

E M

AN

UFA

CTU

RIN

G

95

6

DISCUSSIONTo date, thresholding is the most commonly used segmentation method in medical AM.

However, accurate bone segmentation often requires manual threshold selection, which

still remains a subjective task. Moreover, recent studies suggest that the majority of

inaccuracies that occur during the medical AM process are introduced during the image

acquisition and image processing phases, rather than during the manufacturing, i.e., the 3D

printing process itself [15]–[17]. Such inaccuracies can markedly influence the resulting

STL model (see Figure 6) and subsequently lead to ill-fitting AM implants [18]. Therefore,

the aim of the present study was to assess the impact of manual and automatic default

threshold selection on the reliability and accuracy of skull STL models.

In the present study, all threshold values selected by the four engineers demonstrated

a good intra-observer reliability (ICC>0.9). Furthermore, the inter-observer reliability was

also good (ICC>0.9), as shown in Table 2. Interestingly, all engineers that were blinded

during the experiment selected threshold values for bone that were very close to the grey

values of soft tissues (Figure 4). This resulted in small disjointed structures in the STL

model (marked red in Figure 7) that represent the transition from bone into soft tissue grey

values. Such disjointed “soft-tissue” structures can be manually removed during STL post-

processing [19]. All engineers purposely selected the “soft tissue” threshold values during

bone segmentation in order to incorporate the maximum number of bone-specific grey

values. These grey values are allocated to voxels that represent different tissues during

the CT image reconstruction process. However, during this process, voxels on the bone-

to-soft tissue boundaries that are partially filled with soft tissue are commonly assigned

a lower grey value than bone. This phenomenon is coined the partial volume effect (PVE)

Figure 5. Geometric variations in mm between the highest and lowest thresholded STL models acquired using four different CT scanners (see also Figure 4).

THE

IMPA

CT O

F MA

NU

AL TH

RE

SHO

LD SE

LEC

TION

IN M

ED

ICA

L AD

DITIV

E M

AN

UFA

CTU

RIN

G

96

6

Figure 6 (A - K). Accuracy of all STL models of the female skull acquired using the lowest (left) and highest (middle) mean threshold value selected by the four engineers and the default threshold value of 500 HU (right). The arrows indicate missing data (C, F) or excessive noise (I) in the default-threshold STL models.

[20]. As a consequence of the PVE, voxels may be erroneously allocated to “soft tissue”

instead of “bone”, resulting in data loss and hence bone voids in the STL model (Figure

6). Therefore, engineers should be aware of this phenomenon since data loss can lead to

large inaccuracies in individualized printed medical constructs [18], [20].

Another major finding in this study was the difference between the MDCT and CBCT

DICOM files that were used to construct STL models (Figure 4). One explanation for this

phenomenon is the inherent difference between these technologies. CBCT technology is

THE

IMPA

CT O

F MA

NU

AL TH

RE

SHO

LD SE

LEC

TION

IN M

ED

ICA

L AD

DITIV

E M

AN

UFA

CTU

RIN

G

97

6

typically more heavily affected by image noise and distortions due to the “cone-beam”

geometry of the X-ray beam [21], [22]. In CBCT, the simultaneously irradiated area is

typically larger than in MDCT technology. This causes increased scatter levels and results

in cupping, reduced contrast, and other scatter-induced artifacts in the reconstructed

image. In addition, CBCT images are more subject to cone beam artifacts due to

the large cone beam angle and the imaging geometry comprising a single focal plane.

The cone beam artifacts result from violating Tuy’s sufficiency condition [23] that requires

that each plane intersecting a region of interest must intersect the focal trajectory, i.e.,

the path defining the radiation source position during the imaging. The embodiments

of cone beam artifacts are dependent on the reconstruction algorithm and the imaging

geometry. Typical cone beam artifacts include the elongation of structures in the axial

direction and negative undershoots at sharp edges in the transaxial planes [24]. In CBCT,

Figure 7. MDCT-derived low-threshold STL model of the female cadaver skull (grey) with disjointed “soft-tissue” structures (red).

THE

IMPA

CT O

F MA

NU

AL TH

RE

SHO

LD SE

LEC

TION

IN M

ED

ICA

L AD

DITIV

E M

AN

UFA

CTU

RIN

G

98

6

the focal trajectory consists of a single planar circle or arc that results in a violation of

Tuy’s sufficiency condition in all regions outside the focal plane. The resulting cone beam

artifacts are more pronounced the further away the region of interest is from the focal

plane. In MDCT, the volume that satisfies Tuy’s sufficiency condition is notably larger due

to the helical nature of the focal trajectory.

The presence of artifacts makes the segmentation and hence the thresholding of bone-

specific grey values in CBCT images more cumbersome [25]. This subsequently leads to

a larger variation in manually selected threshold values for CBCT images (Figure 4) and to

the larger geometric variations of up to 0.55 mm in CBCT-derived STL models observed

in this study (Figure 5). DECT-derived STL models demonstrated geometric variations of

up to 0.59 mm (Figure 5). As a consequence of these geometric variations in STL models,

the use of DECT and CBCT technology in its present form does not deliver reproducible

STL models for medical AM. Therefore, the authors of this study suggest that only MDCT

technology should be used for AM applications because of the lower variability (0.13 mm,

see Figure 5) and higher accuracy (Figure 6) of the technology.

The present study demonstrates that the “human factor”, i.e., the medical engineer,

influences the outcome of the segmentation process. Moreover, no single bone threshold

value is applicable for all facial bones. The authors of this study therefore recommend

the use of individual threshold values for each anatomical buttress. Recently, attempts

have been made to develop novel segmentation algorithms using multi-thresholding

[26], adaptive thresholding [11], and semi-automatic region growing [27]. However, these

algorithms are still in an early stage of development [28] and do not take the inherent

differences between MDCT and CBCT technologies into account. Future research should

therefore focus on developing novel medical image segmentation software that is suitable

for different CT imaging modalities. Furthermore, new approaches should be developed

using pattern recognition and machine learning algorithms.

CONCLUSIONThis study shows that manual threshold selection results in better skull STL models than

default thresholding since all the medical engineers in our study selected grey values

closer to soft tissue to compensate for bone voids. Our study also showed that MDCT-

derived STL models offer the lowest variability and highest accuracy, whilst the use of

DECT and CBCT technology in its present form does not deliver reliable STL models

for medical AM. New approaches based on pattern recognition and machine learning

algorithms are required.

STATEMENTSConflicts of interestAuthor 2 Juha Koivisto and author 3 Kalle Karhu are currently employed by Planmeca

Ltd (Finland), a company that specializes in the manufacture of cone-beam computed

tomography scanners. The other authors declare that they have no conflict of interest.

THE

IMPA

CT O

F MA

NU

AL TH

RE

SHO

LD SE

LEC

TION

IN M

ED

ICA

L AD

DITIV

E M

AN

UFA

CTU

RIN

G

99

6

Ethical approvalThis article does not contain any studies with human participants or animals performed

by any of the authors. All human cadaveric materials that were used in the present study

(one female and one male head) were anonymously acquired through the body donor

program of the Department of Anatomy of the VU University Medical Center Amsterdam,

The Netherlands, in full accordance with Article 1 of the Dutch law on funeral services (http://

wetten.overheid.nl/BWBR0005009/2015-07-01) and European legislation. Furthermore,

ethical approval for this study was provided by the Medical Ethical Committee (METC) of

the VU University Medical Center (Ref. 2016.401).

Informed consentFor this type of study, no formal consent was required.

THE

IMPA

CT O

F MA

NU

AL TH

RE

SHO

LD SE

LEC

TION

IN M

ED

ICA

L AD

DITIV

E M

AN

UFA

CTU

RIN

G

100

6

REFERENCES

1. Abou-El Fetouh A, Barakat A, and Abdel-Ghany K (2011) Computer-guided rapid-prototyped templates for segmental mandibular osteotomies: a preliminary report. Int. J. Med. Robot. 7:187–192

2. Johnson T R C (2012) Dual-energy CT: general principles. AJR. Am. J. Roentgenol. 199:3–8

3. Pauwels R, Nackaerts O, Bellaiche N, Stamatakis H, Tsiklakis K, Walker A, Bosmans H, Bogaerts R, Jacobs R, and Horner K (2013) Variability of dental cone beam CT grey values for density estimations. Br. J. Radiol. 86:20120135

4. Razi T, Niknami M, and Alavi Ghazani F (2014) Relationship between Hounsfield Unit in CT Scan and Gray Scale in CBCT. J. Dent. Res. Dent. Clin. Dent. Prospects 8:107–110

5. Katsumata A, Hirukawa A, Okumura S, Naitoh M, Fujishita M, Ariji E, and Langlais R P (2007) Effects of image artifacts on gray-value density in limited-volume cone-beam computerized tomography. Oral Surg. Oral Med. Oral Pathol. Oral Radiol. Endod. 104:829–36

6. Silva I M de C C, Freitas D Q de, Ambrosano G M, Bóscolo F N, and Almeida S M (2012) Bone density: comparative evaluation of Hounsfield units in multislice and cone-beam computed tomography. Braz. Oral Res. 26:550–556

7. Nackaerts O, Maes F, Yan H, Couto Souza P, Pauwels R, and Jacobs R (2011) Analysis of intensity variability in multislice and cone beam computed tomography. Clin. Oral Implants Res. 22:873–9

8. Reeves T, Mah P, and McDavid W (2012) Deriving Hounsfield units using grey levels in cone beam CT: a clinical application. Dentomaxillofacial Radiol. 41:500–508

9. Sahoo PK, Soltani S, and Wong A K C (1988) A survey of thresholding techniques. Comput. Vision, Graph. Image Process. 41:233–260

10. Logan H, Wolfaardt J, Boulanger P, Hodgetts B, and Seikaly H (2013) Evaluation of the accuracy of Cone Beam Computerized Tomography (CBCT): medical imaging technology in head and neck reconstruction. J. Otolaryngol. - Head Neck Surg. 42:25

11. Nackaerts O, Depypere M, Zhang G, Vandenberghe B, Maes F, and Jacobs R (2015) Segmentation of Trabecular Jaw Bone on Cone Beam CT Datasets. Clin. Implant Dent. Relat. Res. 17:1082–1091

12. van Dam A, van Munsteren C, and de Ruiter M (2015) Fix for Life. The Development of a New Embalming Method to Preserve Life-like Morphology. FASEB J. 29:10

13. Bartko J J (1966) The intraclass correlation coefficient as a measure of reliability. Psychol. Rep. 19:3–11

14. Van den Broeck J, Vereecke E, Wirix-Speetjens R, and Van der Sloten J (2014) Segmentation accuracy of long bones. Med. Eng. Phys. 36:949–953

15. Huotilainen E, Paloheimo M, Salmi M, Paloheimo K-S, Björkstrand R, Tuomi J, Markkola A, and Mäkitie A (2014) Imaging requirements for medical applications of additive manufacturing. Acta radiol. 55:78–85

16. Gibson I, Rosen D, and Stucker B (2014) Additive manufacturing technologies - 3D Printing, Rapid Prototyping, and Direct Digital Manufacturing. Springer, New York

17. Smith E J, Anstey J A, Venne G, and Ellis R E (2013) Using additive manufacturing in accuracy evaluation of reconstructions from computed tomography. Proc Inst Mech Eng H 227:551–559

18. Stoor P, Suomalainen A, Lindqvist C, Mesimaki K, Danielsson D, Westermark A, and Kontio R K (2014) Rapid prototyped patient specific implants for reconstruction of orbital wall defects. J. cranio-maxillo-facial Surg. 42:1644–1649

THE

IMPA

CT O

F MA

NU

AL TH

RE

SHO

LD SE

LEC

TION

IN M

ED

ICA

L AD

DITIV

E M

AN

UFA

CTU

RIN

G

101

6

19. Mitsouras D, Liacouras P, Imanzadeh A, Giannopoulos A A, Cai T, Kumamaru K K, George E, Wake N, Caterson E J, Pomahac B, Ho V B, Grant G T, and Rybicki F J (2015) Medical 3D Printing for the Radiologist. Radiographics 35:1965–1988

20. Huotilainen E, Jaanimets R, Valášek J, Marcián P, Salmi M, Tuomi J, Mäkitie A, and Wolff J (2014) Inaccuracies in additive manufactured medical skull models caused by the DICOM to STL conversion process. J. Cranio-Maxillofacial Surg. 42:e259–e265

21. Feldkamp L A, Davis L C, and Kress J W (1984) Practical cone-beam algorithm. J. Opt. Soc. Am. 1:612

22. Endo M, Tsunoo T, Nakamori N, and Yoshida K (2001) Effect of scattered radiation on image noise in cone beam CT. Med. Phys. 28:469-474

23. Tuy H K (1983) An Inversion Formula for Cone-Beam Reconstruction. SIAM J. Appl. Math. 43:546–552

24. Clack R, Zeng G L, Weng Y, Christian P E, and Gullberg G T (1991) Cone beam single photon emission computed tomography using two orbits. In: Colchester

A C F, Hawkes D J (eds) Proceedings of the Information processing in medical imaging: 12th international conference, IPMI ’91 Wye, July 7–12, 1991. Springer, Berlin, pp 45–54

25. Maret D, Telmon N, Peters O A, Lepage B, Treil J, Inglèse J M, Peyre A, Kahn J L, and Sixou M (2012) Effect of voxel size on the accuracy of 3D reconstructions with cone beam CT. Dentomaxillofacial Radiol. 41:649–655

26. Rathnayaka K, Sahama T, Schuetz M A, and Schmutz B (2011) Effects of CT image segmentation methods on the accuracy of long bone 3D reconstructions. Med. Eng. Phys. 33:226–33

27. Xi T, Schreurs R, Heerink W J, Bergé S J, and Maal T J J (2014) A Novel Region-Growing Based Semi-Automatic Segmentation Protocol for Three-Dimensional Condylar Reconstruction Using Cone Beam Computed Tomography (CBCT). PLoS One 9:e111126

28. Pauwels R, Araki K, Siewerdsen J H, and Thongvigitmanee S S (2015) Technical aspects of dental CBCT: state of the art. Dentomaxillofacial Radiol. 44:20140224

7 ACCURACY OF MDCT AND CBCT IN THREE-DIMENSIONAL

EVALUATION OF THE UPPER AIRWAY MORPHOLOGY

Maureen van Eijnatten*, Hui Chen*,

Ghizlane Aarab, Tymour Forouzanfar, Jan de Lange,

Paul van der Stelt, Frank Lobbezoo, Jan Wolff

European Journal of Orthodontics (2017). cjx030 [epub ahead of print]doi: 10.1093/ejo/cjx030

* Both authors contributed equally to this work

AC

CU

RA

CY O

F MD

CT A

ND

CB

CT

104

7

ABSTRACTObjectivesTo assess the accuracy of five different computed tomography (CT) scanners for

the evaluation of the oropharynx morphology

MethodsAn existing cone-beam computed-tomography (CBCT) dataset was used to fabricate an

anthropomorphic phantom of the upper airway volume that extended from the uvula to

the epiglottis (oropharynx) with known dimensions (gold standard). This phantom was

scanned using two multi-detector row computed-tomography (MDCT) scanners (GE

Discovery CT750 HD, Siemens Somatom Sensation) and three CBCT scanners (NewTom

5G, 3D Accuitomo 170, Vatech PaX Zenith 3D). All CT images were segmented by two

observers and converted into standard tessellation language (STL) models. The volume

and the cross-sectional area of the oropharynx were measured on the acquired STL models.

Finally, all STL models were registered and compared with the gold standard.

ResultsThe intra- and inter-observer reliability of the oropharynx segmentation was fair to excellent.

The most accurate volume measurements were acquired using the Siemens MDCT (98.4%;

14.3 cm3) and Vatech CBCT (98.9%; 14.4 cm3) scanners. The GE MDCT, NewTom 5G CBCT

and Accuitomo CBCT scanners resulted in smaller volumes, viz., 92.1% (13.4 cm3), 91.5%

(13.3 cm3), and 94.6% (13.8 cm3), respectively. The most accurate cross-sectional area

measurements were acquired using the Siemens MDCT (94.6%; 282.4 mm2), Accuitomo

CBCT 95.1% (283.8 mm2), and Vatech CBCT 95.3% (284.5 mm2) scanners. The GE MDCT

and NewTom 5G CBCT scanners resulted in smaller areas, viz., 89.3% (266.5 mm2) and

89.8% (268.0 mm2), respectively.

LimitationsImages of the phantom were acquired using the vendor-supplied default airway scanning

protocol for each scanner.

ConclusionSignificant differences were observed in the volume and cross-sectional area measurements

of the oropharynx acquired using different MDCT and CBCT scanners. The Siemens MDCT

and the Vatech CBCT scanners were more accurate than the GE MDCT, NewTom 5G,

and Accuitomo CBCT scanners. In clinical settings, CBCT scanners offer an alternative to

MDCT scanners in the assessment of the oropharynx morphology.

AC

CU

RA

CY O

F MD

CT A

ND

CB

CT

105

7

INTRODUCTIONObstructive sleep apnea (OSA) is a sleep-related breathing disorder, often associated

with a compromised upper-airway space and an increase in upper-airway collapsibility

[1]. The most common complaints of OSA patients are excessive daytime sleepiness,

unrefreshing sleep, poor concentration, and fatigue [2]. OSA also has a range of

deleterious consequences that include increased cardiovascular morbidity, neurocognitive

impairment, and overall mortality [3]–[6]. An important role in the pathogenesis of OSA is

played by anatomical and functional abnormalities of the upper airway [7].

The three-dimensional (3D) morphology of the upper airway is currently assessed using

computed tomography (CT) technologies [8]. The two most common CT technologies

used to date for the assessment of the upper airway are multi-detector row computed

tomography (MDCT) [9] and cone-beam computed tomography (CBCT) [10]. The major

advantages of CBCT scanners are their lower radiation dose and costs [11], [12]. As

a result, CBCT scanners are being increasingly used for upper-airway imaging in OSA

patients [13]–[15]. CBCT scanners use a single, partial gantry rotation [16], which not only

accounts for lower radiation dose, but also produces acceptable diagnostic image quality

[17]. However, it remains unclear whether MDCT and CBCT scanners can provide accurate

3D images of the upper airway.

According to a systematic review by Alsufyani et al. [18], only one out of sixteen studies

focusing on the use of CBCT to automatically or semi-automatically model the upper airway

had a sufficiently sound methodology to test the accuracy of the upper airway dimensions.

The main challenge faced in the assessment of the upper airway accuracy using MDCT or

CBCT is the lack of a “gold standard” model with known dimensions. Recent studies have

used artificial models, hence phantoms of the upper airway, as a gold standard [19]–[21].

However, commercially available phantoms used in most of the aforementioned studies are

commonly manufactured in simple, generic forms and sizes and therefore do not resemble

the clinical situation. In the present study, a novel 3D printed anthropomorphic phantom

of the upper airway volume (oropharynx) with known dimensions was manufactured that

closely resembled a real patient in terms of size, shape, structure, and attenuation profiles.

The aim of this study was to assess the accuracy of two different MDCT scanners and

three different CBCT scanners using a novel 3D printed anthropomorphic phantom for

the evaluation of the oropharynx morphology.

MATERIALS AND METHODSA CBCT dataset of a 27-year-old female that had been previously acquired using

a NewTom 5G CBCT scanner (QR systems, Verona, Italy), was used to design and 3D print

an anthropomorphic phantom of the airway space (Figure 1 A and B). The aforementioned

CBCT dataset was converted into a virtual 3D surface, hence standard tessellation

language (STL) model of the upper airway volume that extended from the uvula to

the epiglottis: the oropharynx (Figure 3). This STL model served as the gold standard

AC

CU

RA

CY O

F MD

CT A

ND

CB

CT

106

7

Figure 1. A. Design of the oropharynx of the phantom (red = maxilla and mandible; green = cervical vertebrae; yellow = supports of the markers; black = markers at the level of the minimum cross-sectional area of the oropharynx; blue = upper airway; purple = base plane; grey and pale-yellow = mould of the skin). B. The 3D printed phantom. C. Sagittal image of the phantom using GE scanner. Arrow = a marker. D. Segmentation based on GE images (purple = oropharynx; green = base plane of the phantom).

in this study. The gold standard STL model of the oropharynx was subsequently used to

manufacture the phantom. All bony structures surrounding the oropharynx were 3D printed

using a High Performance Composite powder ZP151 (3D Systems, Rock Hill, USA). This

composite material was chosen due to its bone-like density that resembles the attenuation

profile of bone [22]. The soft tissue surrounding the oropharynx was fabricated using soft-

tissue-equivalent silicon (Dragon Skin 30, Smooth-On, Inc., Macungie, Pennsylvania, USA).

During the assembling of the phantom, three metal markers were positioned in a defined

plane to acquire a reproducible reference-point system (RPS) for the cross-sectional area

measurement of the oropharynx (Figure 1).

As Figure 2 shows, MDCT images of the oropharynx phantom were acquired using two

MDCT scanners: GE Discovery CT750 HD 64-slice MDCT (GE Healthcare, Little Chalfont,

Buckinghamshire, UK) and Siemens Somatom Sensation 64-slice MDCT (Siemens Medical

AC

CU

RA

CY O

F MD

CT A

ND

CB

CT

107

7

Solutions, Malvern PA, USA). Furthermore, the following three CBCT scanners were used

to acquire CBCT images of the phantom: NewTom 5G CBCT (QR systems, Verona, Italy),

3D Accuitomo 170 CBCT (J Morita, Kyoto, Japan), and Vatech PaX Zenith 3D CBCT

(Vatech, Fort Lee, USA). All MDCT and CBCT images were acquired using the vendor-

supplied default airway scanning protocol. All imaging parameters of the five CT scanners

are presented in Table 1.

The acquired CT datasets were saved as Digital Imaging and Communications in

Medicine (DICOM) files and were imported into Amira® software (v4.1, Visage Imaging

Inc., Carlsbad, CA, USA) (Figure 2C). Using thresholding, one maxillofacial radiologist and

one orthodontist then segmented all the acquired DICOM datasets of the oropharynx

Figure 2. Flowchart of this study.

Table 1. Image acquisition parameters for MDCT and CBCT scans.

GE MDCT

Siemens MDCT

NewTom 5G CBCT

Accuitomo CBCT

Vatech CBCT

Tube voltage (kV) 120 120 110 90 115Tube current (mA) 103 57 5.8 5 6Scan time (s) 0.5 0.5 3.6 17.5 24Slices thickness (mm) 0.625 0.600 0.300 1 0.200Number of voxels 512 x 512x

180512 x 512 x151

610 x 610 x 539

684 x 684 x 480

800 x 800 x 632

Reconstruction Soft B40f Standard N.A. N.A.DLP (mGy-cm)

DAP (mGy-cm2)

CTDIvol (mGy)

46.49

N.A.

N.A.

49

N.A.

N.A.

N.A.

12.232

N.A.

N.A.

N.A.

8.70

N.A.

17.67

N.A.

CBCT: cone beam computed tomography; CTDI: computed tomography dose index; DAP: dose-area product; DLP: dose-length product; Gy: Gray; MDCT: multi-detector row computed tomography; N.A.: not applicable.

AC

CU

RA

CY O

F MD

CT A

ND

CB

CT

108

7

(Table 2). Both observers were blinded for their own results and those of each other.

The segmentation procedure was performed five times for each CT scanner and was

subsequently repeated after a ten-day interval. This resulted in a total of 20 threshold

values per CT scanner (Table 2). These 20 threshold values per scanner were subsequently

used to segment the oropharynx. All segmented oropharynx volumes acquired from

the five MDCT and CBCT scanners (Figure 2D) were converted into STL models. These STL

models were subsequently imported into GOM Inspect v8® metrology software (GOM,

Braunschweig, Germany) for further airway analysis.

The region of interest (ROI) in this study was the volume of the oropharynx between two

parallel planes located 1.5 mm and 42 mm above the base plane of the phantom (Figure 3).

In order to obtain comparable oropharynx volume measurements, all STL models derived

from the five CT datasets were cropped accordingly. The volume of the oropharynx was

subsequently calculated using GOM Inspect software (Figure 4A). Furthermore, the cross-

sectional area of the oropharynx at the level of the metal markers in the phantom was

calculated (Figure 4B).

To determine the accuracy of the CT-derived STL models, all acquired STL models were

superimposed onto the gold standard STL model of the oropharynx using a verified surface

registration (local best-fit) algorithm in GOM Inspect software with an accuracy of 0.05 mm

[23]. All geometric deviations between the oropharynx STL models and the printed gold

standard phantom STL model are depicted in Figure 5.

Figure 3. The segmented volume of the oropharynx. (Red line = upper boundary of region of interest (ROI); yellow line = lower boundary of ROI; green line = the location of the markers)

AC

CU

RA

CY O

F MD

CT A

ND

CB

CT

109

7

Finally, statistical analysis was performed, using IBM Statistical Package for Social

Sciences for Windows (SPSS® version 21, Chicago, Il, USA). Statistical significance was

set at α=0.05. To determine the intra- and inter-observer reliability of the oropharynx

measurements, intraclass correlation coefficients (ICCs) were calculated. Reliability was

divided into three categories: poor (ICC<0.40); fair to good (0.40≤ICC≤0.75); excellent

(ICC>0.75) [24]. The accuracy of the scanners was calculated as the ratio of the phantom

measurements to the gold standard (%). One-way analysis of variance (ANOVA) was used

to test the differences in the volume and cross-sectional area measurements between

the five different CT scanners. Post-hoc analysis (Tukey’s honest significant difference test)

was run to establish which CT scanners produced significantly different results.

RESULTSAll threshold values used for the segmentation of the oropharynx are shown in table 2.

Intra- and inter-observer reliability hence ICCs of the threshold values ranged from 0.436

(fair to good) to 0.966 (excellent).

There were significant differences between the volume measurements of the oropharynx

STL models acquired using the five different CT scanners (F=84.21; P=0.00) (Figure 4A).

Tukey’s test showed that there were no significant differences in volume measurements

between the Siemens MDCT and Vatech CBCT scanners, and between the GE MDCT and

NewTom 5G CBCT scanners. The Siemens MDCT and Vatech CBCT scanners provided

the most accurate volume measurements of the oropharynx (Figure 4A). The NewTom

5G CBCT, Accuitomo CBCT, and the GE MDCT scanners resulted in smaller volume

measurements of the oropharynx (Figure 4A, Table 3).

There were also significant differences between the cross-sectional area measurements

of the oropharynx STL models acquired using the five different CT scanners (F=43.11;

P=0.00) (Figure 4B). The Siemens MDCT, Vatech CBCT and Accuitomo CBCT scanners

Table 2. Intraobserver and interobserver reliability of the threshold values (Hounsfield Units) chosen for five CT scanners estimated by intraclass correlation coefficients (ICCs), based on 20 measurements in total per scanner.

Scanner

Threshold values (HU) (Intra-observer reliability) 

Threshold values (HU)(Inter-observer reliability) 

Observer 1Mean±SD (ICC)

Observer 2Mean±SD (ICC)  Mean±SD (ICC)

GE (MDCT) -339±47 (.966) -389±87 (.864) -364±73 (.818)Siemens (MDCT) -204±54 (.573) -250±77 (.748) -231±70 (.558)NewTom5G (CBCT) -114±35(.481) -102±40 (.720) -108±37 (.786)Accuitomo (CBCT) -289±40 (.661) -244±62 (.438) -267±56 (.507)Vatech (CBCT) -361±75 (.787) -330±63 (.436) -346±70 (.592)

AC

CU

RA

CY O

F MD

CT A

ND

CB

CT

110

7

A.

B.

Figure 4. A. Mean and standard deviation of the volume of the oropharynx derived from five CT scanners. GS = gold standard; NS = no significant difference. B. Mean and standard deviation of the cross-sectional area of the oropharynx derived from five CT scanners. GS = gold standard; NS = no significant difference. The accuracy (%) was calculated as the ratio between the phantom measurements and the gold standard values.

Table 3. Mean and standard deviation of the volume and the cross-sectional area of the upper airway derived from five CT scanners. GS = gold standard.

Variable GSGEMDCT

SiemensMDCT

NewTom 5GCBCT

AccuitomoCBCT

VatechCBCT

Volume of the upper airway (cm3)

14.5 13.4 ± 0.32 14.3 ± 0.31 13.3 ± 0.15 13.8 ± 0.21 14.4 ± 0.19

Area of the upper airway (mm2)

295.5 266.5 ± 6.1 282.4 ± 6.8 268.0 ± 3.8 283.8 ± 7.9 284.5 ± 5.2

provided the most accurate cross-sectional area measurements of the oropharynx (Figure

4B). The GE MDCT and NewTom 5G CBCT scanners resulted in smaller area measurements

of the oropharynx (Figure 4B, Table 3).

Figure 5 shows the oropharynx STL models acquired using five different CT scanners.

The largest geometric deviations of were observed in the vicinity of the uvula and

the epiglottis region (Figure 5).

AC

CU

RA

CY O

F MD

CT A

ND

CB

CT

111

7

DISCUSSION3D evaluation of the oropharynx offers new possibilities of assessing anatomical

abnormalities in OSA patients. In this study, significant differences (P < 0.001) were found

between the volume and cross-sectional area measurements of the oropharynx acquired

using different MDCT and CBCT scanners (Figure 4, Figure 5).

The most accurate volume measurements of the oropharynx were acquired using

the Siemens MDCT (98.4%; 14.3 cm3) and Vatech CBCT (98.9%; 14.4 cm3) scanners

Figure 5. The results of registration between all CT-derived oropharynx STL models (acquired using the mean threshold value) and the gold standard STL model. A = GE; B = Siemens; C = NewTom 5G; D = Accuitomo; E = Vatech; F = Gold standard STL model. Scale: Red = CT-derived STL model is larger than the gold standard; blue = CT-derived STL model is smaller than the gold standard; green = CT-derived STL model is close to the gold standard; black and white = outliers (> 0.8 mm).

AC

CU

RA

CY O

F MD

CT A

ND

CB

CT

112

7

(Figure 4 A). The GE MDCT, NewTom 5G CBCT and Accuitomo CBCT scanners resulted

in smaller volume measurements, viz., 92.1% (13.4 cm3), 91.5% (13.3 cm3), and 94.6%

(13.8 cm3), respectively. The most accurate cross-sectional area measurements of

the oropharynx were acquired using the Siemens MDCT (94.6%; 282.4 mm3), Accuitomo

CBCT (95.1%; 283.8 mm3) and Vatech CBCT (95.3%; 284.5 mm3) scanners (Figure 4 B).

The GE MDCT and NewTom 5G CBCT scanners resulted in smaller area measurements,

viz., 89.3% (266.5 mm3) and 89.8% (268.0 mm3), respectively. These results are in good

agreement with previous studies that reported an underestimation of the airway area in

both MDCT and CBCT images [25], [26]. However, it should be noted that the absolute

values of the aforementioned inaccuracies ranged between 13.3 cm3 and 14.4 cm3 (volume),

and between 266.5 mm2 and 284.5 mm2 (cross-sectional area) (Table 3). Consequently,

the authors of this study hypothesize that the reported inaccuracies should not affect

the radiological evaluation of OSA patients in clinical settings [27].

Even though the authors of this study used the recommended soft tissue imaging

protocols on all CT scanners, the STL models acquired using the GE and NewTom 5G

scanners were generally smaller in size than the gold standard STL model (Figure 4).

The smaller STL models caused a shift in the histograms towards the negative direction

(Figure 5 A, C). This phenomenon is probably due to the partial volume effect [28], in

which voxels in the vicinity of the air-to-soft tissue boundary are commonly allocated to

“soft tissue” instead of “air” during the segmentation process.

The higher accuracy of the Vatech STL models could be a result of the smaller spatial

resolution used in the default airway scanning protocol (Table 1) [29]. However, a very

recent study by Sang et al. [30] that investigated the influence of voxel size on the accuracy

of NewTom 5G and Vatech CBCT reported that increasing the voxel resolution from 0.30

to 0.15 mm does not always result in increased accuracy of 3D tooth reconstructions, while

different CBCT modalities (i.e. NewTom 5G vs. Vatech) can significantly affect the accuracy.

The largest geometric deviations were found in the uvula and epiglottis area (Figure 5).

Interestingly, the acquired oropharynx STL models were generally too large in the epiglottis

region and too small in the vicinity of the uvula. One explanation for this phenomenon

could be that the epiglottis has a concave-like geometry and the uvula is convex. These

findings are in good agreement with a previous study by Barone et al. [31] who observed

discrepancies between the segmentation of concave and convex shapes in teeth.

The results of the present study show that CBCT scanners offer an alternative to MDCT

scanners in the assessment of the oropharynx. This is in good agreement with a previous

study by Suomalainen et al. [32], who reported that CBCT scanners offer images similar to

those acquired using low-dose MDCT protocols. Therefore, taking the lower CBCT radiation

dose into consideration [11], [12], clinicians should preferably use CBCT modalities for

the analysis of the oropharynx. Moreover, all appropriate measures should be undertaken

to minimize the dose according to the International Commission on Radiological Protection

(ICRP) and the As Low As Reasonably Achievable (ALARA) principles [33].

AC

CU

RA

CY O

F MD

CT A

ND

CB

CT

113

7

Limitations of the current studyOne limitation of this study was that for each of the five CT scanners, only one image

dataset was acquired using a single image acquisition protocol. Since there are multiple

image acquisition protocols available for each MDCT and CBCT scanner, different protocols

should be considered in a future study. Another limitation was that the gold standard

values in the present study were obtained from the original STL model of the oropharynx

that was used to 3D print the phantom. 3D printing was performed using a Zprinter 250

inkjet powder printer (3D Systems, Rock Hill, USA), which has a layer thickness of 0.1 mm

and an in-plane resolution of approximately 0.05 mm [23]. Therefore, this process may

have introduced a manufacturing error, hence measurement uncertainty, of up to 0.2 mm

[34]. Nevertheless, this uncertainty can be considered clinically insignificant [35].

CONCLUSION Significant differences were observed in the volume and cross-sectional area measurements

of the oropharynx acquired using different MDCT and CBCT scanners. The Siemens MDCT

and the Vatech CBCT scanners were more accurate than the GE MDCT, NewTom 5G,

and Accuitomo CBCT scanners. In clinical settings, CBCT scanners offer an alternative to

MDCT scanners in the assessment of the oropharynx morphology.

AC

CU

RA

CY O

F MD

CT A

ND

CB

CT

114

7

REFERENCES

1. Ryan C M and Bradley T D (2005) Pathogenesis of obstructive sleep apnea. J. Appl. Physiol. 99:2440–2450

2. Aarab G, Lobbezoo F, Hamburger H L, and Naeije M (2011) Oral appliance therapy versus nasal continuous positive airway pressure in obstructive sleep apnea: a randomized, placebo-controlled trial. Respiration. 81:411–419

3. Marin J M, Carrizo S J, Vicente E, and Agusti A G N (2005) Long-term cardiovascular outcomes in men with obstructive sleep apnoea-hypopnoea with or without treatment with continuous positive airway pressure: an observational study. The Lancet 365:1046–1053

4. McNicholas W T, Strohl K P, White D, Levy P, Schmidt W, Wheatley J, Redline S, Carley D, Buysse D, Young T, Douglas N, and Fleetham J (1999) Sleep-Related Breathing Disorders in Adults: Recommendations for Syndrome Definition and Measurement Techniques in Clinical Research. Sleep 22:667–689

5. Parish J M and Somers V K (2004) Obstructive sleep apnea and cardiovascular disease. In Mayo Clinic Proceedings, 79:1036–1046

6. Villaneuva A T C, Buchanan P R, Yee B J, and Grunstein R R (2005) Ethnicity and obstructive sleep apnoea. Sleep Med. Rev. 9:419–436

7. Lee R W W, Sutherland K, and Cistulli P A (2010) Craniofacial morphology in obstructive sleep apnea: a review. Clin. Pulm. Med. 17:189–195

8. Slaats M A, Van Hoorenbeeck K, Van Eyck A, Vos W G, De Backer J W, Boudewyns A, De Backer W, and Verhulst S L (2015) Upper airway imaging in pediatric obstructive sleep apnea syndrome. Sleep Med. Rev. 21:59–71

9. Abramson Z, Susarla S, August M, Troulis M, and Kaban L (2010) Three-dimensional computed tomographic analysis of airway anatomy in patients

with obstructive sleep apnea. J. oral Maxillofac. Surg. 68:354–362

10. Ogawa T, Enciso R, Shintaku W H, and Clark G T (2007) Evaluation of Cross-section Airway Configuration of Obstructive Sleep Apnea. Oral Surg. Oral Med. Oral Pathol. Oral Radiol. Endod. 103:102–108

11. Ludlow J B and Ivanovic M (2008) Comparative dosimetry of dental CBCT devices and 64-slice CT for oral and maxillofacial radiology. Oral Surgery, Oral Med. Oral Pathol. Oral Radiol. Endodontology 106:106–114

12. Pauwels R, Beinsberger J, Collaert B, Theodorakou C, Rogers J, Walker A, Cockmartin L, Bosmans H, Jacobs R, Bogaerts R, and Horner K (2012) Effective dose range for dental cone beam computed tomography scanners. Eur. J. Radiol. 81:267–271

13. Schendel S A, Broujerdi J A, and Jacobson R L (2014) Three-dimensional upper-airway changes with maxillomandibular advancement for obstructive sleep apnea treatment. Am. J. Orthod. Dentofac. Orthop. 146:385–393

14. Marcussen L, Henriksen J E, and Thygesen T (2017) Do Mandibular Advancement Devices Influence Patients’ Snoring and Obstructive Sleep Apnea? A Cone-Beam Computed Tomography Analysis of the Upper Airway Volume. J. Oral Maxillofac. Surg. 73:1816–1826

15. Cossellu G, Biagi R, Sarcina M, Mortellaro C, and Farronato G (2015) Three-dimensional evaluation of upper airway in patients with obstructive sleep apnea syndrome during oral appliance therapy. J. Craniofac. Surg. 26:745–748

16. Scarfe W C and Farman A G (2008) What is cone-beam CT and how does it work? Dent. Clin. North Am. 52:707–730

17. Lofthag-Hansen S, Thilander-Klang A, and Grondahl K (2011) Evaluation of subjective image quality in relation to

AC

CU

RA

CY O

F MD

CT A

ND

CB

CT

115

7

diagnostic task for cone beam computed tomography with different fields of view. Eur. J. Radiol. 80:483–488

18. Alsufyani N A, Flores-Mir C, and Major P W (2012) Three-dimensional segmentation of the upper airway using cone beam CT: a systematic review. Dentomaxillofacial Radiol. 41:276–284

19. Weissheimer A, de Menezes L M, Sameshima G T, Enciso R, Pham J, and Grauer D (2012) Imaging software accuracy for 3-dimensional analysis of the upper airway. Am. J. Orthod. Dentofac. Orthop. 142:801–813

20. Ghoneima A and Kula K (2013) Accuracy and reliability of cone-beam computed tomography for airway volume analysis. Eur. J. Orthod. 35:256–261

21. Schendel S A and Hatcher D (2010) Automated 3-dimensional airway analysis from cone-beam computed tomography data. J. oral Maxillofac. Surg. 68:696–701

22. Emadi N, Safi Y, Akbarzadeh Bagheban A, and Asgary S (2014) Comparison of CT-Number and Gray Scale Value of Different Dental Materials and Hard Tissues in CT and CBCT. Iran. Endod. J. 9:283–286

23. Mandal D and Syan C (eds) (2016) CAD/CAM, Robotics and factories of the future. Springer, New Delhi, India

24. Fleiss J L (1999) Design and analysis of clinical experiments. John Wiley & Sons, NY

25. Baumgaertel S, Palomo J M, Palomo L, and Hans M G (2009) Reliability and accuracy of cone-beam computed tomography dental measurements. Am. J. Orthod. Dentofac. Orthop. 136:19–25; discussion 25-28

26. Rodriguez A, Ranallo F N, Judy P F, Gierada D S, and Fain S B (2014) CT reconstruction techniques for improved accuracy of lung CT airway measurement. Med. Phys. 41:111911

27. Buchanan A, Cohen R, Looney S, Kalathingal S, and De Rossi S (2016) Cone-beam CT

analysis of patients with obstructive sleep apnea compared to normal controls. Imaging Sci. Dent. 46:9–16

28. Pham D L, Xu C, and Prince J L (2000) Current methods in medical image segmentation. Annu. Rev. Biomed. Eng. 2:315–337

29. Taft R M, Kondor S, and Grant G T (2011) Accuracy of rapid prototype models for head and neck reconstruction. J. Prosthet. Dent. 106:399–408

30. Sang Y-H, Hu H-C, Lu S-H, Wu Y-W, Li W-R, and Tang Z-H (2016) Accuracy Assessment of Three-dimensional Surface Reconstructions of In vivo Teeth from Cone-beam Computed Tomography. Chin. Med. J. (Engl). 129:1464–1470

31. Barone S, Paoli A, and Razionale A V (2016) CT segmentation of dental shapes by anatomy-driven reformation imaging and B-spline modelling. Int. J. Numer. Method. Biomed. Eng. 32:e02747

32. Suomalainen A, Kiljunen T, Kaser Y, Peltola J, and Kortesniemi M (2009) Dosimetry and image quality of four dental cone beam computed tomography scanners compared with multislice computed tomography scanners. Dentomaxillofac. Radiol. 38:367–378

33. International Commission on Radiological Protection (2007) The 2007 Recommendations of the International Commission on Radiological Protection. Elsevier, Amsterdam

34. Lemu H G and Kurtovic S (2012) 3D Printing for Rapid Manufacturing: Study of Dimensional and Geometrical Accuracy. IFIP Adv. Inf. Commun. Technol. 384:470–479

35. Khalil W, EzEldeen M, Van De Casteele E, Shaheen E, Sun Y, Shahbazian M, Olszewski R, Politis C, and Jacobs R (2016) Validation of cone beam computed tomography-based tooth printing using different three-dimensional printing technologies. Oral Surg. Oral Med. Oral Pathol. Oral Radiol. 121:307–315

8 THE ACCURACY OF ULTRASHORT ECHO TIME MRI SEQUENCES FOR

MEDICAL ADDITIVE MANUFACTURING

Maureen van Eijnatten, Erik-Jan Rijkhorst, Mark Hofman,

Tymour Forouzanfar, Jan Wolff

Dentomaxillofacial Radiology (2016). 45(5): 20150424doi: 10.1259/dmfr.20150424

THE

AC

CU

RA

CY O

F ULTR

ASH

OR

T EC

HO

TIME

MR

I SEQ

UE

NC

ES

118

8

ABSTRACTObjectivesAdditively manufactured bone models, implants and drill guides are becoming

increasingly popular amongst maxillofacial surgeons and dentists. To date, such constructs

are commonly manufactured using computed tomography (CT) technology that induces

ionizing radiation. Recently, ultrashort echo time (UTE) magnetic resonance imaging (MRI)

sequences have been developed that allow radiation-free imaging of facial bones. The aim

of the present study was to assess the feasibility of UTE MRI sequences for medical additive

manufacturing (AM).

MethodsThree morphologically different dry human mandibles were scanned using a CT and MRI

scanner. Additionally, optical scans of all three mandibles were made to acquire a “gold

standard”. All CT and MRI scans were converted into Standard Tessellation Language (STL)

models and geometrically compared to the gold standard. To quantify the accuracy of

the AM process, the CT, MRI and gold standard STL models of one of the mandibles were

additively manufactured, optically scanned and compared to the original gold standard

STL model.

ResultsGeometric differences between all three CT-derived STL models and the gold standard

were less than 1.0 mm. All three MRI-derived STL models generally presented deviations

below 1.5 mm in the symphyseal and mandibular area. The AM process introduced minor

deviations of less than 0.5 mm.

ConclusionsThis study demonstrates that MR imaging using UTE sequences is a feasible alternative to

CT in generating STL models of the mandible, and would therefore be suitable for surgical

planning and AM. Further in-vivo studies are necessary to assess the usability of UTE MRI

sequences in clinical settings.

THE

AC

CU

RA

CY O

F ULTR

ASH

OR

T EC

HO

TIME

MR

I SEQ

UE

NC

ES

119

8

INTRODUCTIONVirtual three-dimensional (3D) surgical planning and additive manufacturing (AM)

technologies are increasingly used in oral and maxillofacial surgery. To date, these

advanced technologies have proven to be invaluable in dental implant surgery [1], in

the restoration of mandibular fractures [2], in tumour resections [3] and in maxillofacial

reconstructions [4], [5]. Moreover, additively manufactured medical constructs such as

drill guides, [6] saw guides [7] or individualized implants [8] offer a predictive, functional

and aesthetic outcome, especially in patients with anatomical limitations, insufficient

mandibular or maxillary bone or poor bone density [9].

Medical AM comprises three basic steps. The first step is the acquisition of high-

resolution medical 3D images that are archived as a digital imaging and communications

in medicine (DICOM) file. The second step is to load the DICOM file into virtual planning

software that commonly converts the DICOM file automatically into a virtual 3D surface

model in the Standard Tessellation Language (STL) file format [10]. The resulting STL

model can then be used to design a medical construct using computer aided design (CAD)

software. The last step is the conversion of the STL model into a g-code that is required

to control the AM process.

Currently, 3D medical images intended for oral and maxillofacial AM are acquired

using multi-detector row computed tomography (MDCT) or cone beam computed

tomography (CBCT) technologies [11]. Both MDCT and CBCT provide high resolution

images of the maxillary and mandibular bone and the surrounding tissues [12]. However,

the aforementioned technologies have one major disadvantage: they expose the patient

to ionizing radiation. The dose levels range between 20 and 400 µSv for CBCT and around

1000 µSv for MDCT modalities, which is about 10 to 50 times more than a conventional

panoramic radiograph (about 20 µSv) [13], [14]. Even though device manufacturers

have been developing “ultra” low-dose CT scanning protocols [15]–[17], any X-ray-

based imaging technology inevitably results in some radiation exposure. Therefore, an

alternative, radiation-free imaging modality is still being sought, especially in applications

that require a large number of follow-up scans [18].

Over the past decade, MRI technology has been discussed as an alternative to CT

technology for implant planning [9], [19], [20]. To date, most conventional MRI sequences

only offer optimal soft tissue images and lack the ability to image bone/air interfaces in

the head area. This is because MR image acquisition is based on water microenvironments

and other sources of protons in the human body. Since cortical bone has a low proton

density and a very short T2 relaxation time of about 1.5 ms [21], conventional spin echo

and gradient echo MRI sequences are too slow to acquire a sufficient signal from bone.

As a result, new ultrashort echo time (UTE) sequences have been recently developed

and successfully used for bone imaging [22], [23]. The unique feature of these novel UTE

sequences is that they sample the MR signal with a minimum echo time due to radial

sampling, and rapidly switch to data recording after the radiofrequency (RF) pulse has

been delivered [23].

THE

AC

CU

RA

CY O

F ULTR

ASH

OR

T EC

HO

TIME

MR

I SEQ

UE

NC

ES

120

8

The MRI of bone using UTE sequences is being applied in several emerging clinical

technologies such as PET-MRI attenuation correction [22], high-intensity focused ultrasound

(HIFU) applications [24], MR-guided surgery [25] and MR-only treatment planning and

guidance in radiotherapy [26]. However, to the best of our knowledge no studies have

been performed on the feasibility of UTE MRI sequences in generating STL models for

AM purposes. In this study, we compare the geometric accuracy of UTE MRI-derived STL

models of three morphologically different human mandibles with CT-derived STL models.

Furthermore, additional geometric inaccuracies introduced during the AM process

are assessed.

MATERIALS AND METHODSThree morphologically different human cadaver mandibles of succumbed Dutch

patients with unknown clinical histories and intact bony structures were obtained from

the Department of Anatomy, VU University Medical Center, Amsterdam, The Netherlands.

The mandibles were boiled for 20 hours and all soft tissues were meticulously

removed manually.

The outline of this study is summarized in Figure 1. The three dry mandibles were

scanned using a CT and an MRI scanner (Step 1). In addition, an optical scan of all three

mandibles was acquired using an optical 3D scanner (Artec Spider™, Artec group, Moscow,

Russia) with a point accuracy of 0.05 mm. The optical scans of the three mandibles were

used as gold standard STL models. The CT and MRI scans were then converted to STL

files (Step 2) and aligned with the gold standard STL models. Subsequently, the geometric

deviations between the CT and MRI-derived STL models and the gold standard models

were calculated. Furthermore, the CT, MRI and gold standard STL model of one mandible

were additively manufactured (Step 3) and scanned again with the optical scanner (Step 4)

in order to map additional geometric deviations introduced during the AM process.

MRIThe mandibular cortical bone has a very short T2* of approximately 1.5 ms, and therefore

conventional spin or gradient echo sequences are too slow to acquire any bone signal. In

UTE imaging, the free induction decay signal is sampled directly with a sub-millisecond

echo time, resulting in a MR signal from the surrounding cortical bone [22], [23].

All three mandibles were scanned using an UTE sequence on a Philips Achieva 3T

MRI scanner (Philips Healthcare, Best, the Netherlands) with an eight-channel head

radiofrequency coil using a UTE single echo sequence. The following scan parameters

were used: repetition time = 4.8 ms, echo time1 = 0.14 ms and flip angle = 15°. A non-

selective hard radiofrequency pulse was used for excitation, followed by tuning the receive

coil. Immediately thereafter, free induction decay sampling was started during ramp-up of

the gradients. This procedure resulted in a time between excitation and readout of 0.14

ms. The 3D k-space was radially sampled and regridded to a Cartesian coordinate system

of isotropic 0.5 x 0.5 x 0.5 mm voxels.

THE

AC

CU

RA

CY O

F ULTR

ASH

OR

T EC

HO

TIME

MR

I SEQ

UE

NC

ES

121

8

CT imagingCT imaging was performed with a Siemens Somatom Sensation 64-slice CT scanner

(Siemens Medical Solutions, Erlangen, Germany) using a low-dose scanning protocol.

The scan parameters were as follows: tube voltage = 80 kV, tube current exposure time

product = 150 mAs, pixel size = 0.3 mm x 0.3 mm, slice thickness = 0.6 mm. The images

were then reconstructed using a sharp bone convolution kernel (H60h).

Image processing and STL deviation analysisAll MRI and CT datasets were saved in DICOM file formats and subjected to the image

processing steps described in Figure 2. All CT datasets were imported into Insight

Segmentation and Registration Toolkit® software (Kitware Inc., Clifton Park, NY) and

manually thresholded using 0 Hounsfield Units (HU). The MRI datasets were thresholded

using 15 arbitrary units (AU) (mandible 1 and 2), and 12 AU (mandible 3) (Figure 2, Step 1).

These threshold values were chosen just above the MRI background noise signal level to

acquire an optimal model of the mandibular bone. Subsequently, all segmented datasets

were converted to STL models (Figure 2, Step 2) and imported into MeshLab software

(Visual Computing Lab, Pisa, Italy). The MRI and CT-derived STL models were aligned with

their corresponding gold standard STL models using an iterative closest point algorithm

(Figure 2, Step 3). The signed STL-to-STL distances, that is geometric deviations, between

all MRI and CT-derived STL models and the gold standard STL model were computed

Figure 1. Outline of the study. STL, Standard Tessellation Language; UTE, ultrashort echo time.

THE

AC

CU

RA

CY O

F ULTR

ASH

OR

T EC

HO

TIME

MR

I SEQ

UE

NC

ES

122

8

using Visualization Toolkit® software (Kitware Inc.) (Figure 2, Step 4) and visualised using

colour maps and histograms (Figure 3).

The 95th percentiles of all STL-to-STL distances were calculated using MatLab software

v.2012, (MathWorks®, Natick, MA). To determine the geometric accuracy in the different

areas of interest for surgical planning, the 95th percentiles were calculated for the whole

mandible as well as for five different anatomical regions: 1) right ramus with condyle;

2) right body; 3) symphyseal area with alveolar ridge; 4) left body; 5) left ramus with

condyle (Table 1).

Additive manufacturingThe CT and MRI-derived STL models of mandible 1 were imported into GOM Inspect®

software (GOM GmbH, Braunschweig, Germany) and smoothed with a surface tolerance of

1.0 mm to correct stair-step artefacts. The smoothed CT-derived STL model, the smoothed

MRI-derived STL model and the gold standard STL model of mandible 1 obtained using

the optical scanner were all printed using a Zprinter 250 inkjet powder printer (3D Systems

Inc., Rock Hill, SC). The resulting additively manufactured models were once again scanned

with an optical scanner, aligned with the gold standard STL model, and subjected to STL

deviation analysis according to Steps 3 and 4 in Figure 2.

RESULTSAll geometric deviations between the optical scan “gold standard” STL models and

the CT-derived STL models were below 1.0 mm (Figure 3 A-C), with 95th percentiles < 0.5

mm (Table 1), with the exception of Mandible 1 (0.71 mm). The MRI-derived STL models

generally presented geometric deviations < 2.0 mm (Figure 3 D-F), with 95th percentiles <

1.5 mm (Table 1), again except for Mandible 1 (1.69 mm).

Figure 4 presents the mean and standard deviation (SD) of the signed geometric

deviations in the five different anatomical regions of all CT and MRI STL models. The mean

value indicates to what extent the STL model was smaller or larger than the gold standard,

and the SD indicates the spread in geometric deviations between the STL model and

Figure 2. Overview of image processing steps and software packages involved in the STL analysis. The Insight Segmentation and Registration Toolkit® software was obtained from Kitware Inc., Clifton Park, NY, the MeshLab software was obtained from Visual Computing Lab, Pisa, Italy, and the Visualization Toolkit® software from Kitware Inc., Clifton Park, NY. DICOM, digital imaging and communications in medicine.

THE

AC

CU

RA

CY O

F ULTR

ASH

OR

T EC

HO

TIME

MR

I SEQ

UE

NC

ES

123

8

the gold standard. The deviations in all CT-derived STL models were uniformly distributed

over the different anatomical regions, whereas all MRI-derived STL models showed

a small spread of deviations in the symphyseal area and a larger spread of deviations in

the mandibular condyles (Figure 4).

Minor additional geometric deviations were reported in the additively manufactured

3D models of Mandible 1 (Figure 5). The additively manufactured gold standard 3D model

demonstrated geometric deviations of up to 0.5 mm (Table 2), especially in thin structures

such as the coronoid process and alveolar ridge. Geometric deviations in the additively

manufactured 3D model obtained using CT data (Figure 5 B) were mostly in the vicinity of

the coronoid process. In the 3D model obtained using MRI data (Figure 5 C), deviations

were observed in the coronoid process, mandibular condyles and alveolar ridge.

Figure 3. (A-F) Deviations between the CT-derived Standard Tessellation Language (STL) models and the gold-standard STL model (a–c), and the MRI-derived STL models and the gold-standard STL model (d–f), shown in colour maps and histograms.

THE

AC

CU

RA

CY O

F ULTR

ASH

OR

T EC

HO

TIME

MR

I SEQ

UE

NC

ES

124

8

Dev

iatio

n(m

m)

1.Ram

us(R

)

2.Bod

y (R)

3.Alve

olar rid

ge

4.Bod

y (L)

5.Ram

us(L)

-1.5

-1.0

-0.5

0.0

0.5

1.0

1.5CT vs. gold std. MRI vs. gold std.

Dev

iatio

n(m

m)

1.Ram

us(R

)

2.Bod

y (R)

3.Alve

olar rid

ge

4.Bod

y (L)

5.Ram

us(L)

-1.5

-1.0

-0.5

0.0

0.5

1.0

1.5CT vs. gold std. MRI vs. gold std.

Dev

iatio

n(m

m)

1.Ram

us(R

)

2.Bod

y (R)

3.Alve

olar rid

ge

4.Bod

y (L)

5.Ram

us(L)

-1.5

-1.0

-0.5

0.0

0.5

1.0

1.5CT vs. gold std. MRI vs. gold std.

Mandible 1 Mandible 2 Mandible 3

Figure 4. Mean ± standard deviation of the geometric deviations of five anatomical regions of the mandible in CT- and MRI-derived Standard Tessellation Language models. Std., standard.

Table 1. The 95th percentiles of the geometric deviations of five anatomical regions of the mandible in CT and MRI-derived Standard Tessellation Language models.

CT MRI

Mandible 1 (mm)

Mandible 2 (mm)

Mandible 3 (mm)

Mandible 1 (mm)

Mandible 2 (mm)

Mandible 3 (mm)

Whole mandible 0.71 0.47 0.40 1.69 1.52 1.25 1. Right ramus with condyle 0.49 0.33 0.42 1.89 1.46 1.55 2. Right body 0.30 0.45 0.43 1.03 1.23 0.96 3. Symphyseal area 0.45 0.56 0.45 1.64 0.96 0.70 4. Left body 0.62 0.56 0.30 1.02 1.30 0.86 5. Left ramus with condyle 0.93 0.45 0.34 2.04 1.91 1.53

Table 1 The 95th percentiles of the geometric deviations of five anatomical regions of the mandible in CT and

MRI-derived Standard Tessellation Language models.

CT MRI

Mandible 1

(mm)

Mandible 2

(mm)

Mandible 3

(mm)

Mandible 1

(mm)

Mandible 2

(mm)

Mandible 3

(mm)

Whole mandible 0.71 0.47 0.40 1.69 1.52 1.25

1. Right ramus with condyle 0.49 0.33 0.42 1.89 1.46 1.55

2. Right body 0.30 0.45 0.43 1.03 1.23 0.96

3. Symphyseal area 0.45 0.56 0.45 1.64 0.96 0.70

4. Left body 0.62 0.56 0.30 1.02 1.30 0.86

5. Left ramus with condyle 0.93 0.45 0.34 2.04 1.91 1.53

Table 2. The 50th, 95th and 99th percentiles of the geometric deviations in additively manufactured three-dimensional models of Mandible 1.

Percentile Gold standard (mm) CT (mm) MRI (mm)

50th 0.11 0.21 0.48 95th 0.46 0.88 1.73 99th 1.12 1.32 2.63

THE

AC

CU

RA

CY O

F ULTR

ASH

OR

T EC

HO

TIME

MR

I SEQ

UE

NC

ES

125

8

Figure 5. (A-C) Deviations between the additively manufactured (AM) gold-standard three-dimensional (3D) model and the gold-standard Standard Tessellation Language (STL) model; the AM CT-derived 3D model and the gold-standard STL model; and the AM MRI-derived 3D model and the gold-standard STL model shown in colour maps and histograms.

THE

AC

CU

RA

CY O

F ULTR

ASH

OR

T EC

HO

TIME

MR

I SEQ

UE

NC

ES

126

8

DISCUSSIONToday, there is an ever increasing need for high-resolution 3D images for use in surgical

planning [27], [28]. In this context, CT and CBCT imaging technologies are commonly used

and subsequently induce harmful radiation to the patient [29]. The cumulative lifetime

radiation dose of the population needs to be minimized by reducing the number of overall

examinations and the dose resulting from each individual exposure [30]. Hence, radiation-

free imaging modalities are still being sought for medical additive manufacturing.

MR and CT imagingThe novel UTE MRI sequence used in this study produced STL models that were

morphologically comparable to those acquired using CT technology (Figure 3). In

the MRI- derived STL models, the 95th percentile was generally found to be < 1.5 mm

(Table 1). The only exception was Mandible 1, which had a 95th percentile of 1.69 mm.

These larger geometric deviations were partially due to the submillimetre-thin alveolar

ridge of the mandible that could not be imaged due to the limited spatial resolution

of the MRI scanner (0.5 mm). Indeed, the 95th percentile of the geometric deviations in

the symphyseal area of mandible 1 was 1.64 mm, compared with 0.96 mm and 0.70 mm for

Mandible 2 and 3, respectively (Table 1). Another anatomical region that showed slightly

larger geometrical deviations in the MRI- derived STL models of all three mandibles were

the condyles (Table 1). This phenomenon was most likely caused by a drop in the UTE

MRI signal below the chosen threshold value at the edge of the condyles. This lower

signal could be due to the higher cortical bone density in the condyle regions. In clinical

settings, the presence of soft tissues would offer the possibility to use more advanced

segmentation methods [22], instead of the thresholding approach used in this study.

In MRI technology, the spatial resolution is dependent on the number of frequency

encoding steps and the size of the field of view. Since the size of the mandible determines

the minimal field of view, the resolution is dependent on the frequency encoding, that

is how often the free induction decay signal is sampled. Therefore, the resolution of

the current UTE MRI sequence could be further improved by increasing the duration of

the scan; however, this would be a disadvantage in a clinical setting.

In the CT-derived STL models, the 95th percentile of the geometric deviations

in the whole mandible was generally found to be < 0.5 mm (Table 1). All geometrical

deviations were < 1.0 mm (Figure 3). The capability of CT to capture small bone structures

is also dependent on the spatial resolution of the scanner (0.3 mm), which is directly linked

to the hardware properties such as the detector configuration of the scanner.

The aforementioned results are in agreement with a previous study by White et al. [31],

who obtained MRI and CT-derived additively manufactured 3D models of 10 ovine knees,

and compared caliper measurements of these models with caliper measurements taken

from the real bony anatomy. They reported a mean deviation of 2.15 ± 2.44 mm in their

MRI models and a mean deviation of 0.61 ± 0.41 mm in their CT models. Lee et al. [32]

THE

AC

CU

RA

CY O

F ULTR

ASH

OR

T EC

HO

TIME

MR

I SEQ

UE

NC

ES

127

8

reported a mean deviation of 0.7 ± 0.1 mm between their MRI and CT-derived STL models

of the femur. A more recent study by Van Den Broeck et al. [33] reported RMS errors of

0.55 mm in their CT-derived STL models of the tibia, and 0.56 mm in their MRI-derived STL

models. The results from our study differ from the ones reported by Lee et al. [32] and Van

Den Broeck et al. [33]. This may have been due to differences in the evaluated anatomical

structures and the measuring protocols. Moreover, the differences may also have been

caused by the different MRI sequences used. White et al. [31], Van Den Broeck et al. [33]

and Lee et al. [32] all used conventional spin or gradient echo sequences that commonly

require more manual segmentation of the bone, which in return can be subjective and very

time-consuming.

The UTE MRI protocol used in this study has certain clinical advantages when compared

to CT technologies. The major advantage of MR imaging is the lack of radiation exposure,

which in return allows for multiple imaging sessions to be undertaken. Furthermore,

the UTE sequence used in this study combined with other sequences available on the same

MRI device allows for hard and soft tissue visualisation, respectively. This can be especially

valuable when treating cancer, ameloblastoma patients and orbital fractures [34], [35].

Additive manufacturingIn order to quantify the accuracy of the AM process, the gold standard STL model of

Mandible 1 was first additively manufactured and then optically scanned and compared to

the original STL model (Figure 5 A). The 95th percentile of geometric deviations introduced

during the AM process using the gold standard STL model was 0.46 mm (Table 2). This

result suggests that only minor deviations were introduced during the AM process.

However, because an optical scanner was used to image the bone, the aforementioned

results are not obtainable in clinical settings.

The 95th percentiles of geometric deviations in the MRI and CT-derived additively

manufactured models were higher than those obtained with the optical scanner: 1.73 mm

and 0.88 mm, respectively (Table 2). These results suggest that both CT and MR imaging

have a greater influence on the accuracy of medical additive manufactured constructs than

the additive manufacturing process itself.

Limitations of this studyThe major limitation of the present study was that mandibles without soft tissues were

used. Therefore, the results of this study are not simply generalizable to clinical conditions.

In clinical settings, a second UTE MRI echo image is required to discriminate bone from air

and soft tissues [22]. Furthermore, motion artefacts due to the relatively long duration of

the UTE sequence (several minutes) and metal artefacts caused by orthodontic appliances

or fillings can affect the image quality in vivo. Further cadaver and patient studies are

recommended in order to assess the feasibility of UTE MRI sequences in clinical settings.

THE

AC

CU

RA

CY O

F ULTR

ASH

OR

T EC

HO

TIME

MR

I SEQ

UE

NC

ES

128

8

CONCLUSIONThis study demonstrates that MR imaging of bone using an UTE sequence is a feasible

alternative to CT imaging in the generation of STL models of mandibles with different

morphologies, and would therefore be suitable for surgical planning and additive

manufacturing. The CT and MRI modalities generally introduced geometrical deviations

in the STL models of < 1.0 mm and < 1.5 mm, respectively. The additive manufacturing

process introduced minor deviations of < 0.5 mm. Further in vivo studies are necessary to

assess the feasibility of the UTE MRI sequence in clinical settings.

THE

AC

CU

RA

CY O

F ULTR

ASH

OR

T EC

HO

TIME

MR

I SEQ

UE

NC

ES

129

8

REFERENCES

1. Ramasamy M, Giri, Raja R, Subramonian, Karthik, and Narendrakumar R (2013) Implant surgical guides: From the past to the present. J. Pharm. Bioallied Sci. 5:S98–S102

2. Voss J O, Varjas V, Raguse J-D, Thieme N, Richards R G, and Kamer L (2016) Computed tomography-based virtual fracture reduction techniques in bimandibular fractures. J. Craniomaxillofac. Surg. 44:177–185

3. Yu H, Wang X, Zhang S, Zhang L, Xin P, and Shen S G (2013) Navigation-guided en bloc resection and defect reconstruction of craniomaxillary bony tumours. Int. J. Oral Maxillofac. Surg. 42:1409–1413

4. Schepers R H, Kraeima J, Vissink A, Lahoda L U, Roodenburg J L, Reintsema H, Raghoebar G M, and Witjes M J (2016) Accuracy of secondary maxillofacial reconstruction with prefabricated fibula grafts using 3D planning and guided reconstruction. J. Cranio-Maxillofacial Surg. 44:392-399

5. Li Y, Jiang Y, Ye B, Hu J, Chen Q, and Zhu S (2016) Treatment of Dentofacial Deformities Secondary to Osteochondroma of the Mandibular Condyle Using Virtual Surgical Planning and 3-Dimensional Printed Surgical Templates. J. Oral Maxillofac. Surg. 74:349–368

6. Dawood A, Marti B M, Sauret-Jackson V, and Darwood A (2015) 3D printing in dentistry. Br Dent J 219:521–529

7. Mazzoni S, Bianchi A, Schiariti G, Badiali G, and Marchetti C (2015) Computer-aided design and computer-aided manufacturing cutting guides and customized titanium plates are useful in upper maxilla waferless repositioning. J. Oral Maxillofac. Surg. 73:701–707

8. Kontio R (2014) Update on mandibular reconstruction: computer-aided design, imaging, stem cells and future

applications. Curr. Opin. Otolaryngol. Head Neck Surg. 22:307–315

9. Aguiar M F, Marques A P, Carvalho A C P, and Cavalcanti M G P (2008) Accuracy of magnetic resonance imaging compared with computed tomography for implant planning. Clin. Oral Implants Res. 19:362–365

10. Huotilainen E, Jaanimets R, Valášek J, Marcián P, Salmi M, Tuomi J, Mäkitie A, and Wolff J (2014) Inaccuracies in additive manufactured medical skull models caused by the DICOM to STL conversion process. J. Cranio-Maxillofacial Surg. 42:e259–e265

11. Huotilainen E, Paloheimo M, Salmi M, Paloheimo K-S, Björkstrand R, Tuomi J, Markkola A, and Mäkitie A (2014) Imaging requirements for medical applications of additive manufacturing. Acta radiol. 55:78–85

12. Rengier F, Mehndiratta A, von Tengg-Kobligk H, Zechmann C M, Unterhinninghofen R, Kauczor H-U, and Giesel F L (2010) 3D printing based on imaging data: review of medical applications. Int. J. Comput. Assist. Radiol. Surg. 5:335–341

13. Pauwels R, Beinsberger J, Collaert B, Theodorakou C, Rogers J, Walker A, Cockmartin L, Bosmans H, Jacobs R, Bogaerts R, and Horner K (2012) Effective dose range for dental cone beam computed tomography scanners. Eur. J. Radiol. 81:267–271

14. Lorenzoni D C, Bolognese A M, Garib D G, Guedes F R, and and Eduardo Franzotti SantAnna (2012) Cone-Beam Computed Tomography and Radiographs in Dentistry: Aspects Related to Radiation Dose. Int. J. Dent. 2012:813768

15. Nauer C B, Rieke A, Zubler C, Candreia C, Arnold A, and Senn P (2011) Low-Dose Temporal Bone CT in Infants and Young

THE

AC

CU

RA

CY O

F ULTR

ASH

OR

T EC

HO

TIME

MR

I SEQ

UE

NC

ES

130

8

Children: Effective Dose and Image Quality. Am. J. Neuroradiol. 32:1375–1380

16. Ono K, Hiraoka T, Ono A, Komatsu E, Shigenaga T, Takaki H, Maeda T, Ogusu H, Yoshida S, Fukushima K, and Kai M (2013) Low-dose CT scan screening for lung cancer: comparison of images and radiation doses between low-dose CT and follow-up standard diagnostic CT. Springerplus 2:393

17. Lutz J, Jager V, and Hempel M J (2007) Delineation of temporal bone anatomy: feasibility of low-dose 64-row CT in regard to image quality. Eur. J. Radiol. 17:2638–2645

18. Singh S, Kalra M K, and Moore M A (2009) Dose reduction and compliance with pediatric CT protocols adapted to patient size, clinical indication, and number of prior studies. Radiology 252:200–208

19. Imamura H, Sato H, Matsuura T, Ishikawa M, and Zeze R (2004) A comparative study of computed tomography and magnetic resonance imaging for the detection of mandibular canals and cross-sectional areas in diagnosis prior to dental implant treatment. Clin. Implant Dent. Relat. Res. 6:75–81

20. Kawai T, Asaumi R, Kagawa T, Inadomi D, Yuasa K, and Yosue T (2014) The observation of the incisive branches of mandibular canal – the comparison between MRI vs. CBCT. Clin. Oral Implants Res. 25:2014

21. Horch R A, Nyman J S, Gochberg D F, Dortch R D, and Does M D (2010) Characterization of 1H NMR signal in human cortical bone for magnetic resonance imaging. Magn. Reson. Med. 64:680–687

22. Keereman V, Fierens Y, Broux T, De Deene Y, Lonneux M, and Vandenberghe S (2010) MRI-based attenuation correction for PET/MRI using ultrashort echo time sequences. J. Nucl. Med. 51:812–8

23. Aitken A P, Giese D, Tsoumpas C, Schleyer P, Kozerke S, Prieto C, and

Schaeffter T (2014) Improved UTE-based attenuation correction for cranial PET-MR using dynamic magnetic field monitoring. Med. Phys. 41:12302

24. Huisman M, Lam M K, Bartels L W, Nijenhuis R J, Moonen C T, Knuttel F M, Verkooijen H M, van Vulpen M, and van den Bosch M A (2014) Feasibility of volumetric MRI-guided high intensity focused ultrasound (MR-HIFU) for painful bone metastases. J. Ther. Ultrasound 2:16

25. Liao H, Inomata T, Sakuma I, and Dohi T (2010) 3-D augmented reality for MRI-guided surgery using integral videography autostereoscopic image overlay. IEEE Trans. Biomed. Eng. 57:1476–1486

26. Metcalfe P, Liney G P, Holloway L, Walker A, Barton M, Delaney G P, Vinod S, and Tomé W (2013) The Potential for an Enhanced Role for MRI in Radiation-Therapy Treatment Planning. Technol. Cancer Res. Treat. 12:429–446

27. Schendel S A and Jacobson R (2009) Three-Dimensional Imaging and Computer Simulation For Office-Based Surgery. J. Oral Maxillofac. Surg. 67:2107–2114

28. Giannatsis J and Dedoussis V (2009) Additive fabrication technologies applied to medicine and health care: a review. Int. J. Adv. Manuf. Technol. 40:116–127

29. Koivisto J (2015) The use of MOSFET dosimeters and anthropomorphic phantoms in low dose dental CBCT applications. Helsinki, Finland: University of Helsinki

30. Implications of commission recommendations that doses be kept as low as readily achievable (1973) ICRP publication 22. Oxford, UK: Pergamon Press

31. White D, Chelule K L, and Seedhom B B (2008) Accuracy of MRI vs CT imaging with particular reference to patient specific templates for total knee replacement surgery. Int. J. Med. Robot. Comput. Assist. Surg. 4:224–231

THE

AC

CU

RA

CY O

F ULTR

ASH

OR

T EC

HO

TIME

MR

I SEQ

UE

NC

ES

131

8

32. Lee Y S, Seon J K, Shin V I, Kim G-H, and Jeon M (2008) Anatomical evaluation of CT-MRI combined femoral model. Biomed. Eng. Online 7:6

33. Van den Broeck J, Vereecke E, Wirix-Speetjens R, and Vander Sloten J (2014) Segmentation accuracy of long bones. Med. Eng. Phys. 36:949–953

34. Apajalahti S, Kelppe J, Kontio R, and Hagström J (2015) Imaging characteristics of ameloblastomas and diagnostic value of computed

tomography and magnetic resonance imaging in a series of 26 patients. Oral Surg. Oral Med. Oral Pathol. Oral Radiol. 120:e118–e130

35. Schmutz B, Rahmel B, McNamara Z, Coulthard A, Schuetz M, and Lynham A (2014) Magnetic resonance imaging: An accurate, radiation-free, alternative to computed tomography for the primary imaging and three-dimensional reconstruction of the bony orbit. J. Oral Maxillofac. Surg. 72:611–618

9 GENERAL DISCUSSION

GE

NE

RA

L DISC

USSIO

N

135

9

Even though surgical techniques have improved over time, surgery still remains a manual

task. Current surgical techniques in the field of oral and maxillofacial surgery are still heavily

dependent on the dexterity and experience of the surgeon. In this context, medical AM

opens new avenues to improve the dexterity of the surgeon during surgical procedures.

The resulting decrease in operation time and improved accuracy can markedly reduce

treatment costs. For example, AM can be combined with virtual planning to create

accurate and low-cost individualized orbital floor implants. This novel treatment method

has shown to reduce operation time by approximately 40% [1]. However, creating accurate

medical AM constructs still requires technical and computational knowledge that can be

a challenge for clinicians. Therefore, AM constructs are currently most often designed

by engineers with in-depth knowledge of the field of anatomical modelling and reverse

engineering.

While the aforementioned AM constructs can be fabricated with great accuracy

(< 0.1 mm) [2] using current printing technologies, their design is currently limited by

the accuracy of the medical images used during the fabrication process. It must be noted

that image quality can differ markedly between MDCT, DECT and CBCT scanners. [3]

Moreover, scanner-specific acquisition parameters can also influence the image quality

[4]. Of all the scanners on the market, MDCT scanners offer the best image quality

and accuracy for medical AM [3]. However, radiation safety still needs to be taken into

consideration when acquiring images for AM, given the current concerns of radiation-

induced secondary malignancies [5]. Therefore, if possible, low-dose MDCT scanning

protocols should be used to minimise the effective dose in patients who require AM

constructs [3]. Even though current CBCT devices can attain lower dose values than MDCT

devices in the maxillofacial area [6], their image quality remains problematic for medical

AM purposes. The CBCT-derived STL models evaluated in this PhD thesis demonstrated

large non-uniform inaccuracies of up to 0.9 mm in the maxillofacial area (Chapter 3), which

was probably due to the relatively high amount of noise and number of artefacts [7] and

inhomogeneities in the CBCT X-ray beam [8]. Bearing the above-mentioned concerns in

mind, more emphasis should be put into developing UTE MRI sequences that are specially

designed for AM purposes [9].

One possible method of improving the overall image quality of CBCT scanners in

the maxillofacial region is to change the head position in the gantry. According to our

studies, the best CBCT image quality can be achieved on a mobile gantry device using prone

or supine imaging positions [10]. More specifically, the prone imaging position resulted in

the best image quality for the sinus, mandible and maxilla when compared with the supine

and oblique imaging positions [11]. The oblique imaging position offered inadequate

image quality except in the sinus region. These novel insights regarding the image quality

of CBCT scanners may help to improve the accuracy of CBCT images for medical AM

purposes. A more advanced method of improving the accuracy of CBCT images would

be to develop adaptive, flexible scan protocols that would allow physicians to zoom into

anatomical areas of interest during image acquisition. For example, an initial “coarse”

GE

NE

RA

L DISC

USSIO

N

136

9

CBCT scan of a patient’s head could be acquired using a large field of view, followed

by a more precise local scan of the region of interest, e.g., a tumour or fracture site. This

process would, however, require real-time reconstruction of CBCT images using powerful

cluster computers. Another promising new technology that could revolutionise the overall

image quality required in the AM process is spectral imaging [12]. Spectral imaging refers

to a technique that uses photon-counting detectors to separate photons coming from

a polychromatic X-ray source into a series of energy bins, for which the energy threshold

levels can be configured using dedicated software. Each tissue type in the human body

has a different response curve to photon energies. Therefore, it is possible to discriminate

tissues based on their response to different photon energies. This could greatly improve

tissue discrimination and delineation in CT and CBCT images.

In addition to the aforementioned CT imaging process, the conversion of the resulting

images into STL models remains a challenge in medical AM. The most challenging step in

the conversion process is image segmentation. Our literature review on bone segmentation

methods used in medical AM (Chapter 6) revealed a large spread in accuracies amongst

the current methods used ranging from 0.04 mm to 1.9 mm. Image segmentation is

often compromised by factors such as the partial volume effect, scattering artefacts and

noise [13].

To date, global thresholding remains the most commonly used segmentation method

in medical AM. However, the selection of a suitable threshold value for bone still remains

a manual task that can introduce geometric variations amongst engineers [14]. Generally,

global thresholding achieves accuracies under 0.6 mm [15]. Advanced thresholding

methods could improve the accuracy of CT-derived STL models to under 0.38 mm [16].

However, such methods are, to the best of our knowledge, not included in the FDA-

approved software packages for medical AM. Finally, it should be noted that the triangle-

based STL file format was not initially developed to describe complex biological structures

that encompass many curved surfaces and different tissue densities. Consequently, many

small triangles are needed to accurately describe a curved surface. A solution could be to

employ novel file formats such as STL2 [17] or the additive manufacturing file format (AMF)

[18]. However, these novel file formats are not yet supported by the software packages

and printers currently used in medical AM.

Current CT and CBCT scanners deliver voxel-based datasets in which anatomical

structures are represented by a continuous spectrum of grey values that are related

to attenuation coefficients. Medical AM, on the other hand, requires binary datasets

in the form of STL models. Therefore, the current medical AM process always requires

a conversion from continuous data into binary data, i.e., image segmentation. A major

drawback with the image segmentation process is that a large amount of the valuable

biological information that was initially present in the CT image in the form of grey values

is discarded. This can be considered as a waste of photons. One solution to this problem

would be to develop an algorithm that could directly compute mesh volumes (STL

models) from raw CT projection data. This could be achieved by restricting the number

GE

NE

RA

L DISC

USSIO

N

137

9

of grey values in a reconstructed CT image to a small, discrete number of values known

in advance: one for each tissue type, i.e., “air”, “soft tissue”, “cancellous bone” and

“cortical bone”. Consequently, since each voxel is already assigned to a specific tissue

type during the reconstruction phase, image segmentation as a whole becomes redundant.

This technique is known as discrete tomography [19]. The major advantage of discrete

tomography over conventional CT is that it maximises the amount of information obtained

from a CT scan for a given dose level: “making the most out of each photon”. Therefore,

I hypothesize that discrete tomography could markedly reduce the number of projections

and radiation dose needed to acquire an accurate STL model.

As mentioned above, the creation of accurate medical AM constructs still requires

a large amount of human intervention by experienced clinicians and engineers, particularly

in the imaging and image processing steps. This problem could be tackled by automating

certain processes in the medical AM workflow using novel artificial intelligence (AI)

approaches [20]. The basic mathematical concept of AI dates back to the 1940s when

researchers proposed the development of an electrical network that mimics the function of

neurons in the human brain [21]. This was the onset for further developments in the field of

AI and led to a variety of machine learning (ML) algorithms that are today used in the fields

of computer vision and medical imaging, among others. In recent years, the development

of graphic processing units (GPUs) has enabled the use of multi-layered ML algorithms

such as convolutional neural networks (CNNs) [22]. Currently, CNNs are becoming

the methodology of choice for analysing medical images [23] and show great potential

in performing accurate and automated image segmentation tasks [24]–[26]. Furthermore,

a CNN can be taught to reconstruct complex 3D shapes from CT data [27] and could

therefore theoretically be used to auto-complete STL models of disfigured skulls. A CNN

can be taught in a supervised or unsupervised manner. To date, the majority of CNNs

in the field of medical imaging use supervised learning, in which a CNN “learns from

experience” based on a large amount of labelled training data (e.g., medical images).

It should be pointed out that the performance of this class of CNN depends heavily on

the availability and the quality of the images used for training. As opposed to supervised

learning, in unsupervised learning a CNN is taught to independently discover patterns

in data. Taking the above into account, one can postulate that recent developments in

AI combined with medical imaging (“big”) data will revolutionise healthcare and will

subsequently offer new possibilities in the field of medical AM, virtual/augmented reality

and robotics. These developments will most likely lead to more automation, which, in

turn, will have a fundamental socio-economic impact on the global labour market in

the 21st century.

GE

NE

RA

L DISC

USSIO

N

138

9

REFERENCES

1. Vehmeijer M, van Eijnatten M, Liberton N, and Wolff J (2016) A Novel Method of Orbital Floor Reconstruction Using Virtual Planning, 3-Dimensional Printing, and Autologous Bone. J. Oral Maxillofac. Surg. 74:1608-1612

2. Pinto J M, Arrieta C, Andia M E, Uribe S, Ramos-Grez J, Vargas A, Irarrazaval P, and Tejos C (2015) Sensitivity analysis of geometric errors in additive manufacturing medical models. Med. Eng. Phys. 37:328–334

3. van Eijnatten M, Berger F H, de Graaf P, Koivisto J, Forouzanfar T, and Wolff J (2017) Influence of CT parameters on STL model accuracy. Rapid Prototyp. J. 23:679–685

4. Huotilainen E, Paloheimo M, Salmi M, Paloheimo K-S, Björkstrand R, Tuomi J, Markkola A, and Mäkitie A (2014) Imaging requirements for medical applications of additive manufacturing. Acta radiol. 55:78–85

5. Mathews J D, Forsythe A V, Brady Z, Butler M W, Goergen S K, Byrnes G B, Giles G G, Wallace A B, Anderson P R, Guiver T A, McGale P, Cain T M, Dowty J G, Bickerstaffe A C, and Darby S C (2013) Cancer risk in 680 000 people exposed to computed tomography scans in childhood or adolescence: data linkage study of 11 million Australians. BMJ 346:f2360

6. Pauwels R, Beinsberger J, Collaert B, Theodorakou C, Rogers J, Walker A, Cockmartin L, Bosmans H, Jacobs R, Bogaerts R, and Horner K (2012) Effective dose range for dental cone beam computed tomography scanners. Eur. J. Radiol. 81:267–271

7. Pauwels R, Nackaerts O, Bellaiche N, Stamatakis H, Tsiklakis K, Walker A, Bosmans H, Bogaerts R, Jacobs R, and Horner K (2013) Variability of dental cone beam CT grey values for density estimations. Br. J. Radiol. 86:20120135

8. Feldkamp L A, Davis L C, and Kress J W (1984) Practical cone-beam algorithm. J. Opt. Soc. Am. 1:612

9. van Eijnatten M, Rijkhorst E-J, Hofman M, Forouzanfar T, and Wolff J (2016) The accuracy of ultrashort echo time MRI sequences for medical additive manufacturing. Dentomaxillofacial Radiol. 45:20150424

10. Koivisto J, van Eijnatten M, Jarnstedt J, Holli-Helenius K, Dastidar P, and Wolff J (2017) Impact of prone, supine and oblique patient positioning on CBCT image quality, contrast-to-noise ratio and figure of merit value in the maxillofacial region. Dentomaxillofac. Radiol. 46:20160418

11. Heiland M, Schulze D, Blake F, and Schmelzle R (2005) Intraoperative imaging of zygomaticomaxillary complex fractures using a 3D C-arm system. Int. J. Oral Maxillofac. Surg. 34:369–375

12. Cormode D P, Si-Mohamed S, Bar-Ness D, Sigovan M, Naha P C, Balegamire J, Lavenne F, Coulon P, Roessl E, Bartels M, Rokni M, Blevis I, Boussel L, and Douek P (2017) Multicolor spectral photon-counting computed tomography: in vivo dual contrast imaging with a high count rate scanner. Scientific Reports, 7:4784

13. Sharma N and Aggarwal L M (2010) Automated medical image segmentation techniques. J. Med. Phys. 35:3–14

14. van Eijnatten M, Koivisto J, Karhu K, Forouzanfar T, and Wolff J (2016) The impact of manual threshold selection in medical additive manufacturing. Int. J. Comput. Assist. Radiol. Surg. 12:607–615

15. Gelaude F, Van der Sloten J, and Lauwers B (2008) Accuracy assessment of CT-based outer surface femur meshes. Comput. Aided Surg. 13:188–199

16. Rathnayaka K, Sahama T, Schuetz M A, and Schmutz B (2011) Effects of CT image segmentation methods on the accuracy

GE

NE

RA

L DISC

USSIO

N

139

9

of long bone 3D reconstructions. Med. Eng. Phys. 33:226–33

17. Hiller J D and Lipson H (2009) STL 2.0: a proposal for a universal multi-material Additive Manufacturing File format. In: Proceedings of the Solid Freeform Fabrication Symposium, 266–278

18. ASTM (2016) Standard Specification for Additive Manufacturing File Format (AMF). Available: http://standards.iso.org/iso/52915

19. Batenburg K J and Sijbers J (2006) Discrete tomography from micro-CT data: application to the mouse trabecular bone structure. In Proc. of SPIE 6142: 614240–614241

20. Lugar G and Stubblefield W (1993) Artificial Intelligence: structures and strategies for complex problem solving. Pearson education

21. McCulloch W S and Pitts W (1943) A logical calculus of the ideas immanent in nervous activity. Bull. Math. Biophys. 5:115–133

22. Hinton G E, Osindero S, and Teh Y-W (2006) A fast learning algorithm for deep belief nets. Neural Comput. 18:1527–1554

23. Litjens G, Kooi T, Bejnordi B E, Setio A A A, Ciompi F, Ghafoorian M, van der Laak J A W M, van Ginneken B, and Sánchez C I (2017) A survey on deep learning in medical image analysis. arXiv Prepr. arXiv1702.05747

24. Krizhevsky A, Sutskever I, and Geoffrey E. H (2012) ImageNet Classification with Deep Convolutional Neural Networks. Adv. Neural Inf. Process. Syst. 25:1–9

25. Prasoon A, Petersen K, Igel C, Lauze F, Dam E, and Nielsen M (2013) Deep feature learning for knee cartilage segmentation using a triplanar convolutional neural network. In: Lecture Notes in Computer Science, 8150 LNCS: 246–253

26. De Brébisson A and Montana G (2015) Deep neural networks for anatomical brain segmentation. In: IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops. 20–28

27. Wu Z, Song S, Khosla A, Yu F, Zhang L, Tang X, and Xiao J (2015) 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 1912–1920

10 ENGLISH SUMMARY & DUTCH SUMMARY

EN

GLISH

SUM

MA

RY & D

UTC

H SU

MM

ARY

143

10

SUMMARYAM offers possibilities to fabricate different types of patient-specific constructs such as

anatomical models, surgical guides and implants. Such constructs can improve the surgeon’s

dexterity, decrease operation time and improve accuracy of surgical procedures. As

demonstrated in Chapter 2, AM can be combined with virtual planning to create accurate

and low-cost individualised orbital floor implants. A virtual 3D model of a patient with

an orbital floor fracture was generated from CT images of the patient. The fracture was

virtually closed using spline interpolation. The resulting 3D model was used to design

and manufacture a mould of the defect site using an inkjet printer. This tangible mould

was subsequently used during surgery to sculpture an individualised autologous orbital

floor implant. This novel treatment method enhanced the overall accuracy and efficiency

of the surgical procedure. The sculptured autologous orbital floor implant showed an

excellent fit in vivo and reduced the operation time by approximately 40%.

Despite recent advances and promising case studies involving medical AM, notable

scientific, technological and regulatory challenges remain. Each step in the medical AM

process, i.e., imaging (step 1), image processing (step 2) and manufacturing (step 3) can

introduce inaccuracies in the resulting construct. However, since the aforementioned

AM constructs can be fabricated with great accuracy (< 0.1 mm) using current printing

technologies, their design is currently limited by the accuracy of the imaging and image

processing steps. Therefore, the general aim of this thesis was to provide an insight into

the current scientific and technological challenges faced in medical additive manufacturing

with respect to imaging and image processing. More specifically, the different parameters

that influence the accuracy of patient-specific AM constructs were identified.

In Chapter 3, the image quality and the accuracy of STL models acquired using different

CT scanners and acquisition parameters is assessed. Images of three dry human skulls

were acquired using different scanning protocols on two multi-detector row computed

tomography (MDCT) scanners, a dual energy computed tomography (DECT) scanner and

one cone beam computed tomography (CBCT) scanner. All images were ranked according

to their image quality and converted into STL models. These STL models were subsequently

compared with corresponding gold standard models acquired using an optical 3D surface

scanner. It was found that the image quality differed between the MDCT, DECT and CBCT

scanners. Images acquired using low-dose MDCT protocols were preferred over images

acquired using routine protocols. All CT-based STL models demonstrated non-uniform

geometrical deviations of up to +0.9 mm. The largest deviations were observed in CBCT-

derived STL models. Therefore, it can be concluded that CT imaging technologies and

their acquisition parameters can markedly affect the accuracy of medical AM constructs.

In Chapter 4, the impact of head positioning on CBCT image quality is evaluated.

The impact of supine, prone and oblique patient imaging positions on the image quality,

contrast-to-noise ratio and figure of merit value in the maxillofacial region of a fresh frozen

cadaver head was assessed using a CBCT scanner. In addition, the CBCT supine images

EN

GLISH

SUM

MA

RY & D

UTC

H SU

MM

ARY

144

10

were compared with supine MSCT images. The best CBCT image quality was achieved

in the prone imaging position for sinus, mandible and maxilla, followed by the supine

and oblique imaging positions. The MSCT scanner offered similar image qualities to

the 7.5-mA supine images acquired using the mobile CBCT scanner. The prone imaging

position offered the best CNR and FOM values on the mobile CBCT scanner. These

findings can help to improve CBCT image quality in the maxillofacial region, which, in

turn, may improve the overall accuracy of CBCT-derived AM constructs.

In addition to the CT imaging step, the accuracy of medical AM constructs is affected

by errors introduced during image processing, particularly during the image segmentation

process. Therefore, in Chapter 5, thirty-two publications that reported on the accuracy of

different CT image segmentation methods for bone used in medical AM are reviewed.

The advantages and disadvantages of the different image segmentation methods used

in these studies were evaluated and the reported accuracies were compared. The spread

between the reported accuracies was large (0.04 mm to 1.9 mm). Global thresholding

was the most commonly used segmentation method with accuracies under 0.6 mm.

The disadvantage of this method is the extensive manual post-processing required.

Advanced thresholding methods could improve the accuracy to under 0.38 mm. However,

such methods are currently not included in commercial software packages. Statistical

shape model methods resulted in accuracies from 0.25 mm to 1.9 mm but they are

only suitable for anatomical structures with moderate anatomical variations. To improve

the accuracy and reduce the cost of patient-specific AM constructs, more advanced image

segmentation methods are required.

Chapter 6 provides more insight into the most commonly used image segmentation

method in medical AM, namely global thresholding. The impact of manual and default

threshold selection on the reliability and accuracy of skull STL models acquired using

different CT technologies was evaluated. One female and one male human cadaver head

were imaged using MDCT, DECT and two CBCT scanners. Four medical engineers manually

thresholded the bony structures on all CT images. The lowest and highest selected mean

threshold values and the default threshold value were used to generate skull STL models.

Geometric variations between all manually thresholded STL models were calculated.

Additionally, in order to calculate the accuracy of the manually and default thresholded STL

models, all STL models were superimposed on an optical scan of the dry female and male

skulls (“gold standard”). The intra- and inter-observer variability of the manual threshold

selection was good (intra-class correlation coefficients >0.9). All engineers selected grey

values closer to soft tissue to compensate for bone voids. Geometric variations between

the manually thresholded STL models were 0.13 mm (MDCT), 0.59 mm (DECT) and 0.55

mm (CBCT). All STL models demonstrated inaccuracies ranging from −0.8 mm to +1.1 mm

(MDCT), −0.7 mm to +2.0 mm (DECT) and −2.3 mm to +4.8 mm (CBCT). The findings of

this study demonstrate that manual threshold selection results in better STL models than

default thresholding.

EN

GLISH

SUM

MA

RY & D

UTC

H SU

MM

ARY

145

10

In Chapter 7, a novel method was developed to assess the accuracy and reliability of CT

scanners and software packages. To this end, an anthropomorphic phantom that provided

ground truth measurements for the evaluation of the oropharynx morphology was 3D

printed. This phantom was used to assess the accuracy of two multi-detector row computed

tomography (MDCT) scanners (GE Discovery CT750 HD, Siemens Somatom Sensation)

and three CBCT scanners (NewTom 5G, 3D Accuitomo 170, Vatech PaX Zenith 3D). All

CT images were segmented by two observers and converted into standard tessellation

language (STL) models. The volume and the cross-sectional area of the oropharynx

were measured on the acquired STL models. Finally, all STL models were registered and

compared with the gold standard. Significant differences were observed in the volume

and cross-sectional area measurements of the oropharynx acquired using the different

MDCT and CBCT scanners. The Siemens MDCT and the Vatech CBCT scanners were more

accurate than the GE MDCT, NewTom 5G, and Accuitomo CBCT scanners.

AM medical models, implants and drill guides are becoming increasingly popular

amongst maxillofacial surgeons and dentists. This rise in popularity has subsequently

increased the number of CT and CBCT scans performed worldwide. However, all X-ray

based imaging modalities induce harmful ionizing radiation to the patient. Therefore,

Chapter 8 assesses the feasibility of using radiation-free UTE MRI sequences for medical

AM. Three morphologically different dry human mandibles were scanned using a CT and

MRI scanner. All CT and MRI scans were converted into STL models and geometrically

compared with a corresponding gold standard STL model acquired using an optical 3D

scanner. To quantify the accuracy of the AM process, the CT, MRI and gold-standard STL

models of one of the mandibles were additively manufactured, optically scanned and

compared with the original gold standard STL model. Geometric differences between all

three CT-derived STL models and the gold standard were < 1.0 mm. All three MRI-derived

STL models generally presented deviations < 1.5 mm in the symphyseal and mandibular

area. It should be noted that the AM process only introduced minor deviations of < 0.5

mm. Hence, UTE MRI sequences offer an alternative to CT in generating STL models of

the mandible and could be suitable for surgical planning and AM.

EN

GLISH

SUM

MA

RY & D

UTC

H SU

MM

ARY

147

10

SAMENVATTINGAdditive manufacturing (AM), oftewel driedimensionaal (3D-) printen, maakt het mogelijk

om patiëntspecifieke producten te fabriceren zoals anatomische modellen, chirurgische

boormallen en implantaten. Dergelijke toepassingen verkorten de operatieduur

en kunnen de nauwkeurigheid van de operatie verbeteren. In Hoofdstuk 2 wordt een

nieuwe behandelmethode gepresenteerd waarbij 3D-printen gecombineerd wordt met

virtuele chirurgische planning om nauwkeurige en goedkope patiëntspecifieke implantaten

te fabriceren voor de reconstructie van de bodem van de oogkas (orbitabodem). De eerste

stap van deze nieuwe behandelmethode is om computertomografie (CT-) beelden van

een patiënt met een orbitabodemfractuur te converteren naar een virtueel 3D-model.

Vervolgens wordt het defect in de orbitabodem virtueel gesloten. Op basis van het

resulterende 3D-model wordt vervolgens een mal ontworpen die geprint wordt met

een inkjet 3D-printer. De chirurg kan een dergelijke fysieke mal tijdens de operatie als

hulpmiddel gebruiken om van autoloog bot een patiëntspecifiek orbitabodemimplantaat

optimaal vorm te geven. Deze nieuwe behandelmethode leidt tot een hogere chirurgische

precisie en efficiëntie en kan de operatietijd significant verkorten.

Ondanks de recente ontwikkelingen en veelbelovende medische casuïstiek op het

gebied van 3D-printen zijn er nog vele wetenschappelijke, technologische en reglementaire

uitdagingen te overwinnen. Het 3D-printen van een medisch product vereist drie stappen:

1) medische beeldvorming; 2) beeldverwerking; en 3) 3D-printen. Elk van deze stappen

kan de nauwkeurigheid van medische 3D-geprinte producten beïnvloeden. Echter, omdat

moderne 3D-printtechnieken zeer nauwkeurig zijn (< 0,1 mm), is de nauwkeurigheid met

name afhankelijk van de medische beeldvorming en de daaropvolgende beeldverwerking.

Het doel van deze dissertatie is inzicht te verkrijgen in de huidige wetenschappelijke

en technologische uitdagingen op het gebied van medisch 3D-printen, met name

omtrent medische beeldvorming en beeldverwerking. Ons onderzoek heeft geleid tot

de identificatie van verschillende parameters die de nauwkeurigheid van medische

3D-geprinte producten beïnvloeden.

In Hoofdstuk 3 wordt de invloed van verschillende CT-scanners en scanparameters

onderzocht op de beeldkwaliteit en de nauwkeurigheid van de resulterende

virtuele 3D-modellen. Drie droge menselijke schedels werden gescand met twee

multidetectorcomputertomografie (MDCT) scanners, een dual-energy computertomografie

(DECT) scanner, en een cone-beam computertomografie (CBCT) scanner.

Op elk van deze CT-scanners werden verschillende scanprotocollen getest. Twee

ervaren radiologen beoordeelden de beeldkwaliteit van de resulterende CT-scans. Hierna

werden de scans geconverteerd naar virtuele 3D-modellen en werd de nauwkeurigheid

van deze 3D-modellen bepaald door ze te vergelijken met een “gouden standaard”

(optische 3D-scan). De radiologen namen significante verschillen in beeldkwaliteit waar

tussen de MDCT-, DECT- en CBCT-scanners. Bovendien prefereerden de radiologen het

lage-dosis-scanprotocol boven het standaard scanprotocol op de MDCT-scanner. Alle

EN

GLISH

SUM

MA

RY & D

UTC

H SU

MM

ARY

148

10

virtuele 3D-modellen vertoonden niet-uniforme onnauwkeurigheden tot +0,9 mm, waarvan

de modellen verkregen met CBCT de grootste afwijkingen vertoonden. De belangrijkste

conclusie van Hoofdstuk 3 is dat het type CT-scanner en het gebruikte scanprotocol een

grote invloed kunnen hebben op de nauwkeurigheid van medische 3D-geprinte producten.

In Hoofdstuk 4 wordt de invloed van de hoofdpositie in de CBCT-scanner op

de beeldkwaliteit, contrast-to-noise ratio (CNR) en figure-of-merit (FOM) onderzocht.

Mobiele CBCT-scanners kunnen het hoofd van een patiënt in rugligging, in buikligging, of

zittend met het hoofd in een schuine positie scannen. In elk van deze drie posities werd een

CBCT-scan gemaakt van een vers ingevroren kadaverhoofd. Ook werd er ter vergelijking

een MDCT scan gemaakt van het kadaverhoofd. De beste beeldkwaliteit voor de

sinussen, mandibula en maxilla op de CBCT-scans werd verkregen in buikligging, gevolgd

door de rugligging. De schuine positie gaf de laagste beeldkwaliteit. De beeldkwaliteit

van de CBCT-scan gemaakt van het kadaverhoofd in rugligging met een buisstroom van

7,5 mA was vergelijkbaar met de MDCT-scan. De rugligging resulteerde in de beste CNR

en FOM voor de CBCT-scanner. Deze nieuwe inzichten over de kwaliteit van CBCT-scans

kunnen de algehele nauwkeurigheid van 3D-geprinte producten verkregen uit CBCT-scans

ten goede komen.

Het is niet alleen de medische beeldvorming die de nauwkeurigheid van medische

3D-printen beïnvloedt maar ook de beeldverwerking. De belangrijkste stap van

beeldverwerking is beeldsegmentatie. Hoofdstuk 5 geeft daarom een overzicht van

tweeëndertig wetenschappelijke artikelen die rapporteren over de nauwkeurigheid van

verschillende CT-beeldsegmentatiemethodes voor botstructuren. De voor- en nadelen

van deze methodes worden samengevat en de nauwkeurigheden worden met elkaar

vergeleken in een tabel. Deze tabel laat zien dat de nauwkeurigheden in de literatuur

sterk uiteenlopen: 0,04 mm tot 1,9 mm. De meestgebruikte segmentatiemethode

in medisch 3D-printen is de zogenaamde drempelmethode (“thresholding”). Deze

methode resulteert in nauwkeurigheden onder 0,6 mm. Een belangrijk nadeel

van de drempelmethode is dat men vaak handmatig ruis of artefacten van het virtuele

3D-model moet verwijderen. Geavanceerde drempelmethodes kunnen de nauwkeurigheid

van virtuele 3D-modellen verbeteren tot onder 0,38 mm. Echter, dergelijke methodes

zijn tot op heden niet beschikbaar in de commerciële softwarepakketten voor medisch

3D-printen. Ten slotte resulteren de segmentatiemethodes gebaseerd op statistische

vormmodellen in nauwkeurigheden tussen 0,25 mm en 1,9 mm, al zijn dergelijke methodes

alleen geschikt voor het segmenteren van botstructuren met beperkte anatomische

variaties. Kortom, betere beeldsegmentatiemethodes zouden de nauwkeurigheid en

de kosten van medisch 3D-printen ten goede komen.

Hoofdstuk 6 gaat dieper in op de drempelmethode. Een medisch ingenieur

selecteert meestal handmatig de juiste grijswaarde als drempelwaarde om botstructuren

te segmenteren in CT-beelden, al bevatten de meeste softwarepakketten ook een

automatische drempelwaarde. Daarom wordt in Hoofdstuk 6 de invloed onderzocht

van handmatige en automatische drempelwaardeselectie op de betrouwbaarheid en

EN

GLISH

SUM

MA

RY & D

UTC

H SU

MM

ARY

149

10

de nauwkeurigheid van virtuele 3D modellen verkregen met verschillende

CT-scanners. Een mannelijk en een vrouwelijk gebalsemd kadaverhoofd werden gescand

met een MDCT-, een DECT- en twee CBCT-scanners. Vier medisch ingenieurs selecteerden

vervolgens in iedere CT-scan een drempelwaarde om bot te segmenteren. De laagste

en de hoogste geselecteerde drempelwaarde werden vervolgens gebruikt om virtuele

3D-modellen te genereren. Dit werd ook gedaan met de automatische drempelwaarde.

De geometrische variaties tussen alle handmatig gegenereerde 3D-modellen werden

berekend. Ook werd de nauwkeurigheid van alle 3D-modellen bepaald door deze

te vergelijken met een gouden standaard (een optische 3D-scan van de schedels

waarvan het zachte weefsel verwijderd was). De intra- en inter-observer variabiliteit van

de handmatige drempelwaardeselectie was goed (intra-class correlatiecoëfficiënt > 0,9).

Alle medisch ingenieurs selecteerden drempelwaardes die leken op de grijswaarden

van de zachte, omliggende weefsels om zoveel mogelijk bot te segmenteren. Echter, er

werden geometrische variaties tussen de resulterende 3D-modellen waargenomen van

0,13 mm (MDCT), 0,59 mm (DECT), en 0,55 mm (CBCT). De nauwkeurigheden liepen

uiteen van -0,8 mm tot +1,1 mm (MDCT), −0,7 mm tot +2,0 mm (DECT), en van −2,3

mm tot +4,8 mm (CBCT). Een andere interessante bevinding was dat de handmatig

geselecteerde drempelwaardes resulteerden in betere 3D-modellen dan die verkregen

met de automatische drempelwaarde.

In Hoofdstuk 7 wordt een nieuwe methode gepresenteerd om de betrouwbaarheid en

de nauwkeurigheid van CT-scanners en softwarepakketten te onderzoeken. Deze methode

bestaat uit het 3D-printen van een antropomorfisch fantoom waarvan de afmetingen

bekend zijn en die dus dienst kan doen als gouden standaard. Een dergelijk fantoom is

in Hoofdstuk 7 gebruikt om de betrouwbaarheid en de nauwkeurigheid van verschillende

CT-scanners te bepalen in het afbeelden van de bovenste luchtwegen (oropharynx). Het

fantoom werd gescand met twee MDCT-scanners (GE Discovery CT750 HD, Siemens

Somatom Sensation) en drie CBCT-scanners (NewTom 5G, 3D Accuitomo 170, Vatech

PaX Zenith 3D). De resulterende CT-scans werden door twee tandartsen gesegmenteerd

en vervolgens geconverteerd naar virtuele 3D-modellen. Van alle 3D-modellen werd het

volume en de doorsnede berekend en vergeleken met de gouden standaard. Er werden

significante verschillen in volume en doorsnede waargenomen tussen de verschillende

MDCT- en CBCT- scanners. De Siemens MDCT- en de Vatech-CBCT scanners waren

nauwkeuriger dan de GE MDCT-, NewTom 5G, en Accuitomo CBCT-scanners.

Medisch 3D-printen wint snel in populariteit onder mond-, kaak- en aangezichtschirurgen

en tandartsen. Dit heeft echter ook tot gevolg dat het aantal gemaakte CT- en CBCT-scans

wereldwijd toeneemt. Alle beeldvormende modaliteiten waarbij röntgenstraling gebruikt

wordt, stellen de patiënt bloot aan schadelijke straling. Daarom wordt in Hoofdstuk 8 een

nieuwe, stralingsvrije methode onderzocht om bot in beeld te brengen: een ultrashort

echo time (UTE) sequentie op een magnetic resonance imaging (MRI-) scanner. Drie

droge mandibulae werden gescand met een UTE MRI-sequentie en een CT-scanner.

De resulterende scans werden geconverteerd naar virtuele 3D-modellen en vergeleken

EN

GLISH

SUM

MA

RY & D

UTC

H SU

MM

ARY

150

10

met een gouden standaard (een optische 3D-scan). Ook werd de nauwkeurigheid van het

3D-printproces onderzocht door alle virtuele 3D-modellen te printen op de inkjet printer

van het VU medisch centrum en deze opnieuw te vergelijken met de gouden standaard.

Alle virtuele 3D-modellen verkregen met de CT-scanner resulteerden in nauwkeurigheden

< 1,0 mm. Ter vergelijking, alle virtuele 3D-modellen verkregen met de UTE MRI-sequentie

waren < 1,5 mm nauwkeurig in de symphysis mandibulae en de corpus mandibulae. Een

andere interessante bevinding was dat het 3D printproces slechts minimale afwijkingen

introduceerde (< 0,5 mm). UTE MRI-sequenties kunnen dus een alternatief bieden voor

CT voor het genereren van virtuele 3D-modellen van de mandibula. Dergelijke modellen

kunnen gebruikt worden voor virtuele chirurgische planning en 3D-printen.

11 ACKNOWLEDGEMENTS

AC

KN

OW

LED

GE

ME

NTS

155

11

ACKNOWLEDGEMENTSThe 3D Innovation Lab of the VU University Medical Center (VUmc) constantly explores and

invests in 3D imaging and printing technologies and a wide array of medical virtual planning

software solutions required for the fabrication of individualized medical constructs hence

human spare parts. This unique collaboration and education space formed the basis of

my PhD research. Therefore, I would like to express my gratitude to all people involved in

the 3D Innovation Lab: Niels Liberton, Sjoerd te Slaa, Frank Verver, Micha Paalman, Dafydd

Visscher, dr. Jack van Loon, Angela Ridwan-Pramana, dr. Sophie Kommers, Omar Hertgers,

Ivar Kommers, Rianne van Loenen, Professor Ruud Verdaasdonk and our colleagues from

the Academic Medical Center (AMC), Professor Theo Smit, dr. Geert Streekstra en dr.

Iwan Dobbe. Also thanks to our Bachelor and Master students that contributed to our

research over the past years: Roelof van Dijk, David Spies, Fedde Kuilman, Niki Tokkari,

Daan Brons, Karen Kiers, Guus van Baar, Janne Ploeger, Jordi Minnema. Meghan Simons,

Sevde Yen, Claire Olthoff, and Ezra Esmeijer. In this context, I would like to thank Theo

Faes for coordinating our Master students.

I would like to thank the department of Oral & Maxillofacial Surgery/Pathology and my

department head and promotor Professor Tymour Forouzanfar for facilitating my PhD. Also

thanks to all of my colleagues from the department. I am particularly grateful to Annelies

van der Geest and dr. Martijn van Steenbergen for their consistent support and excellent

organization skills.

Dr. Eric-Jan Rijkhorst and dr. Dennis Heijtel, I would like to thank you for the interesting

physics-related discussions and your contributions to our research. Eliane Kaaij and

Jasmine Rubirayoxall, thank you for your outstanding anatomical skills and passion. And

Professor Paul van der Stelt and dr. Hui Chen, thank you for the productive collaboration.

A special thanks to Planmeca Oy for providing financial support for my PhD project. Timo

Müller, Vesa Varjonen, Jaakko Lähelmä, dr. Kalle Karhu, dr. Ari Hietanen, and dr. Juha

Koivisto, I hope we can continue our excellent collaboration in the future. I would also

like to offer my thanks to DSM Biomedical B.V., and particularly Jac Koenen, for their

financial support. Finally, I would like to thank Anke Kleinveld and her colleagues from

the Institute for Education and Training for allowing me to become a tutor at the VUmc

School of Medical Sciences and thereby supporting my research at the VUmc. It was

an amazing experience for me to guide Bachelor students through their first years of

medical education.

Thanks so much to my family and friends for their unlimited encouragement and support. In

this context, I would like to thank my paranymphs Elise van Eijnatten and Mike van Rijssel.

Floor Maarse, thank you for being such a great swim teacher. Furthermore, I am especially

AC

KN

OW

LED

GE

ME

NTS

156

11

grateful to Daniël Doorman, for always taking the time to listen to me and standing by my

side. A special thanks to Peter Heath for revising the manuscript and to my dad for revising

the Dutch summary. I am also grateful to all of the co-authors of the related publications.

Finally, I am extremely grateful to my mentor and co-promotor Jan Wolff for believing in

me from the beginning of my research career and for his consistent support.