Reduce in Variation and Improve Efficiency of Target Volume Delineation by a Computer-Assisted System Using a Deformable Image Registration Approach

Int. J. Radiation Oncology Biol. Phys., Vol. 68, No. 5, pp. 1512–1521, 2007Copyright � 2007 Elsevier Inc.

Printed in the USA. All rights reserved0360-3016/07/$–see front matter

doi:10.1016/j.ijrobp.2007.04.037

PHYSICS CONTRIBUTION

REDUCE IN VARIATION AND IMPROVE EFFICIENCY OF TARGET VOLUMEDELINEATION BY A COMPUTER-ASSISTED SYSTEM USING A DEFORMABLE

IMAGE REGISTRATION APPROACH

K. S. CLIFFORD CHAO, M.D.,* SHREERANG BHIDE, FRCR,y HANSEN CHEN, M.S.,z JOSHUA ASPER, PAC,#

STEVEN BUSH, M.D.,x GREGG FRANKLIN, M.D., PH.D.,x VIVEK KAVADI, M.D.,k

VICHAIVOOD LIENGSWANGWONG, M.D.,{ WILLIAM GORDON, M.D.,# ADAM RABEN, M.D.,z

JON STRASSER, M.D.,z CHRISTOPHER KOPROWSKI, M.D.,z STEVEN FRANK, M.D.,*

GREGORY CHRONOWSKI, M.D.,* ANESA AHAMAD, M.D.,* ROBERT MALYAPA, M.D., PH.D.,**

LIFEI ZHANG, PH.D.,yy AND LEI DONG, PH.D.yy

*Department of Radiation Oncology, the University of Texas M. D. Anderson Cancer Center, Houston, TX; yDepartment of Head &Neck Oncology, Royal Marsden Hospital, London, United Kingdom; zDepartment of Radiation Oncology, Christiana Care Health

Services, Newark, DE; # IKOE Education and Training Unit, Houston, TX; xDepartment of Radiation Oncology, New Mexico CancerCenter, Albuquerque, NM; kDepartment of Radiation Oncology, Texas Oncology Cancer Center, Sugar Land, TX; {Department ofRadiation Oncology, Minnesota Oncology Hematology, Maplewood, MN; # Department of Radiation Oncology, Texas Oncology,

Webster, TX; **Department of Radiation Oncology, the University of Florida Proton Treatment Center, Jacksonville, FL; andyyDepartment of Radiation Physics, the University of Texas M. D. Anderson Cancer Center, Houston, TX

Purpose: To determine whether a computer-assisted target volume delineation (CAT) system using a deformableimage registration approach can reduce the variation of target delineation among physicians with different headand neck (HN) IMRT experiences and reduce the time spent on the contouring process.Materials and Methods: We developed a deformable image registration method for mapping contours from a tem-plate case to a patient case with a similar tumor manifestation but different body configuration. Eight radiationoncologists with varying levels of clinical experience in HN IMRT performed target delineation on two HN cases,one with base-of-tongue (BOT) cancer and another with nasopharyngeal cancer (NPC), by first contouring fromscratch and then by modifying the contours deformed by the CAT system. The gross target volumes were provided.Regions of interest for comparison included the clinical target volumes (CTVs) and normal organs. The volumetricand geometric variation of these regions of interest and the time spent on contouring were analyzed.Results: We found that the variation in delineating CTVs from scratch among the physicians was significant, andthat using the CAT system reduced volumetric variation and improved geometric consistency in both BOT andNPC cases. The average timesaving when using the CAT system was 26% to 29% for more experienced physiciansand 38% to 47% for the less experienced ones.Conclusions: A computer-assisted target volume delineation approach, using a deformable image-registrationmethod with template contours, was able to reduce the variation among physicians with different experiencesin HN IMRT while saving contouring time. � 2007 Elsevier Inc.

Target delineation, Deformable image registration, Contour variations, Auto-segmentation, Computer-assisted.

INTRODUCTION

Technologic advances in radiation therapy have now made

intensity-modulated radiation therapy (IMRT) a feasible

option in routine clinical practice. The dosimetric superiority

of IMRT offers the advantage of better tumor target coverage

151

and critical normal tissue sparing and has generated

promising treatment results, particularly in patients with

prostate and head-and-neck (HN) cancers (1–6). Key to the

successful implementation of IMRT is the proper delineation

of target volumes on planning computed tomography (CT)

Reprint requests to: K.S. Clifford Chao, M.D., Department of Ra-diation Oncology, Unit 97, The University of Texas M. D. AndersonCancer Center, 1515 Holcombe Blvd., Houston, TX 77030.Tel: (713) 563-2300; Fax: (713) 563-2331; E-mail: [email protected]

Conflict of interest: This research study was conducted and com-pleted before the execution of a license agreement between the Uni-

versity of Texas M.D. Anderson Cancer Center and IKOEtech,LLC, to further develop this technology for clinical use. K.S. Clif-ford Chao is a co-founder of IKOEtech, LLC.

Received Jan 3, 2007, and in revised form March 15, 2007.Accepted for publication April 6, 2007.

2

mailto:[email protected]

mailto:[email protected]

Computer-assisted contouring based on deformable image registration d K. S. C. CHAO et al. 1513

images. Furthermore, implementation of IMRT requires

additional resources to be allocated so that physicians have

a block of time to perform contouring. Miles et al. (7) high-

lighted the logistics involved in resource allocation for differ-

ent steps of IMRT implementation and found that the time

spent by physicians on a HN IMRT case was, on average,

2.3 h, whereas Hong et al. (8) observed significant variation

in target volume determination (what to include) and delinea-

tion (where to draw) among physicians from 20 esteemed in-

stitutions. Consistently contouring the clinical target volume

(CTV) in HN cancer indeed poses a great challenge for

physicians with varying training backgrounds and levels of

clinical experience in HN IMRT.

To address this challenge and in an attempt to improve the

consistency and efficiency of target volume delineation, we

developed a computer-assisted target volume delineation

(CAT) system that implements a deformable image registra-

tion method to morph a disease-specific contour template to

the patient’s CT scans. The main objectives of this study were

to assess whether CAT could reduce contouring variation

among physicians with different HN IMRT experience and

to determine whether the time spent on the laborious contour-

ing process could be shortened.

METHODS AND MATERIALS

IMRT experience of participating physiciansEight volunteer board-certified radiation oncologists participated

in this contouring study. Because we used HN examples in this

study, the physicians’ backgrounds in practice and experience

with HN IMRT are pertinent to the study objectives and are shown

in Table 1. Four of the oncologists, who have annual HN IMRT

experience of 20 cases or fewer, were grouped as less experienced

or ‘‘sporadic’’ in the timesaving study analysis; the remaining

four physicians, who have more experience (i.e., more than 20 cases

annually), were grouped as ‘‘frequent.’’

Study proceduresAll participating physicians were asked to contour on images

from the same two patient imaging sets, one with a T1N2a base-

of-tongue (BOT) cancer and one with a T1N1 nasopharyngeal can-

cer (NPC). Participants were provided a brief clinical history, the

staging information, and both primary and nodal gross tumor vol-

umes (GTV). Participating physicians were asked to contour each

image set twice, once using a familiar commercial planning system

for contouring from scratch (CFS) and once with the assistance of

our CAT system using template contours that were generated by

an experienced radiation oncologist (K.S.C.C.) based on the clinical

scenarios, the joint RTOG/EORTC/DAHANCA guidelines, and

publications from reputable institutions on target volume definition

in HN cancers (4, 9–14). These template contour sets (one for BOT

and one for NPC cases) were meant as a reference and allowed for

modification by the participating physicians based on the following

instructions on CTV and normal tissue contouring. Considering that

physicians might be influenced by the template contours if they per-

form CAT-assisted contouring first followed by CFS, the current

practice patterns (CFS) in HN target delineation of all participating

physicians were captured first before exposing them to the template

contours. Participating physicians performed CAT-assisted contour-

ing immediately after CFS.

Contouring guidelinesThe CTV represents the volume to be treated, and it is further

classified as follows: the CTV1 denotes the volume that encom-

passes the GTV with a 1-cm margin. The CTV is truncated by bones

and air cavities that are not involved by the tumors in these two test

cases. The CTV2 includes additional margin beyond that of the

CTV1 and the ipsilateral nodal regions immediately adjacent to

the GTV of the primary tumor site or in the neck that are deemed

to be at high risk of containing microscopic tumor. The CTV3 in-

cludes low-risk regions, such as the uninvolved contralateral neck,

but that require prophylactic irradiation. The participating physi-

cians were asked to contour the following normal organs at risk

(OARs): left and right parotid glands, spinal cord, brain, brainstem,

and larynx. For the patient with the NPC, the orbits, optic nerves,

and optic chiasm were also contoured. Mistakes in determining

and delineating GTV will lead to faulty CTVs; however, defining

GTV requires physicians to integrate information from clinical man-

ifestations, physical examination, and diagnostic images which were

not feasible and available in the current study setting. To eliminate

the variation in defining GTV, GTVs for both cases were provided

when participating physicians contoured either from scratch (CFS)

or with assistance from by CAT.

CAT apparatusOur CAT system enables physicians to apply contour templates to

their patients’ CT images. The CAT procedure is carried out through

two phases of image registration: the rigid bony registration and the

deformable image registration. Both registration algorithms are fully

automatic. The process constitutes a few simple steps. First, the phy-

sician selects the study patient. Second, the physician chooses the

matching template as predetermined in the study design. The CAT

image interface then performs anatomic landmark recognition and

contours the template’s translational alignment (Fig. 1a), followed by

image and contour deformation onto the slices of the patient’s CT

images (Fig. 1b). Third, the physician then adjusts the CTV and nor-

mal tissue contours on the CT images of the patient case with side-

by-side and slice-by-slice reference to the template case (Fig. 1c).

After contour modification is complete, the contour data are exported

for analysis. The total deformation times for the patient cases were 68

s for the BOT cancer and 107 s for the NPC. These times were added

to that of the total CAT process in data analysis.

The image deformation and registration software used in this

study was developed in house and is an enhanced form of the

‘‘Demons’’ algorithm developed by Thirion (15). To improve the

Table 1. Details of physicians’ levels of experience

Physicianno.

Radiationoncology

experience (y)

Averagepatientstreated

per year

AverageIMRT

patientsper year

Average HNIMRT

patientsper year

1 23 120 60 42 17 307 32 73 13 220 45 104 1 300 50 205 32 300 60 256 1 403 49 277 15 233 100 678 12 402 232 91Median 14 300 55 23

Abbreviations: HN = head and neck; IMRT = intensity-modu-lated radiation therapy.

1514 I. J. Radiation Oncology d Biology d Physics Volume 68, Number 5, 2007

Fig. 1. Schematic illustration of a deformable imaging-registration method for computer-assisted contouring. The imageinterface performs anatomic landmark recognition and then aligns and overlays the patient’s computed tomography (CT)images with the template images (a), followed by contour deformation on the patient’s images (b). The physician then ad-justs the clinical target volume and normal tissue contours on the patient’s CT images with side-by-side and slice-by-slicereference to the template contours (c).

convergence and the searching range, we combined the Demons

algorithm with a multi-resolution approach: we modified the Demon

forces in the original algorithm to include an ‘‘active’’ force in addi-

tion to the ‘‘passive’’ force used by the original algorithm. This

modification was made to improve the convergence and shorten

the processing time. The enhanced algorithm has been validated

in a study and described in a peer-reviewed publication (16).

Data acquisition and analysisThe CFS procedure was performed by the physicians who chose

a familiar treatment planning system among two commercial treat-

ment planning systems: the Eclipse (Varian Medical Systems,

Palo Alto, CA) and the Pinnacle3 (Philips Medical Systems, Bothell,

WA). The contours created with the CAT software after deformable

image registration and physician contour modification were


exported using the DICOM-RT format to the Pinnacle3 treatment

planning system for data analysis. Similarly, the contours manually

drawn in the Eclipse treatment planning were exported to the Pinna-

cle3 treatment planning system for data analysis. We did this be-

cause using the same treatment planning to report volumes of the

contoured structures will minimize any differences between two

treatment planning systems. Descriptive statistics (average, standard

deviations, etc.) for these contoured volumes were calculated. Re-

duction in contour volume variation was quantified by comparing

the standard deviations of contour volumes generated by either

CFS or CAT.

To determine contour volume agreement with and without using

the CAT, we computed a volume overlap index (VOI), which is

defined by the following equation:

VOI ¼ VcontouredXVreferenced�Vcontoured þ Vreferenced

��2

in which Vcontoured is the volume of the manually drawn (CFS) con-

tours or the CAT-generated and physician modified contours and

Vreferenced is the CAT-generated reference contours. The union of

Vcontoured and Vreferenced represents the overlapped area of the two

volumes. Ideally, if the contours from scratch match the CAT-gen-

erated reference contours perfectly, the value of the VOI is 1. Indi-

vidual VOI values were computed for various structures (CTVs and

normal organs at risk) contoured by the individual physicians.

In addition to the VOI to measure the volume overlap, we calcu-

lated the surface disagreement between the two contoured volumes.

We used the Euclidean distance transformation to measure the mean

distance between the two contoured surfaces in three dimensions

(17).

TimesavingWe measured both the time (in min) for each physician to perform

CFS and to use the CAT software and then calculated the difference

between the times. In addition, we calculated the ratio of the time

required to use CAT to the time required for CFS. As described

above, for the timesaving analysis, we arbitrarily categorized the

physicians into two groups based on the frequency of their HN

IMRT.

Statistical analysisThe results (variation in target volumes, VOI, three-dimensional

surface disagreement, and time spent) from the two groups, i.e., CFS

and CAT, were compared using the paired t test and Wilcoxon’s test

for paired nonparametric data. Statistical analysis was performed us-

ing Statistical Program for Social Sciences (SPSS Inc., Chicago, IL)

software.

RESULTS

Variation in target volumesFor the case of the patient with BOT cancer, the standard

deviations of the volumes of CTV1, CTV2, and CTV3 were

25.7, 65.6, and 113.7 when using CFS, and 6.0, 7.1, and 7.2

when using CAT, respectively (Table 2). For the case of the

patient with NPC, the standard deviations of the volumes of

CTV1, CTV2, CTV3 were 83.9, 88.9, and 65.6 when using

CFS, and 7.9, 18.3, and 9.9 when using CAT, respectively

(Table 3). We observed significant variation in the contoured

CTV volumes between the eight physicians’ when they

performed CFS. We found that most participating physicians

drew smaller CTV1 and CTV2 for BOT than those in the

templates and in some cases and did not comply with the min-

imum 1 cm margin as stated in the study instruction (Table 2

and Figure 2). However, this discrepancy largely reduced by

using CAT-assisted approach. The CTV1 of CFS for the

NPC case (Table 3) was larger because one single physician

drew a 330-cc CTV1 as compared with approximately 100 cc

in the template. Removing this outlier data point would re-

duce the average CTV1 for CFS to 114 cc (compared with

141 cc) and the standard deviation to 38 cc (as compared

with �84 cc). The variation was higher in the case of NPC

in part due to the complexity of anatomical structures near the

base of skull, and also because the standard deviation was

significantly reduced when the CAT system was used. This

phenomenon was evident in both the BOT and NPC cases

Table 2. Contouring variation among participatingphysicians in the case of the patient with base-of-tongue cancer

CFS CAT-assistedTemplatecontours

ROI

Averagevolume(cm3) 1 SD


Volume(cm3)

CTV1 80.9 25.7 143.97 5.96 141.33CTV2 226.3 65.6 296.11 7.13 304.34CTV3 144.5 113.7 117.79 7.24 119.93Left parotid 30.6 5.6 25.61 2.06 24.92Right parotid 23.1 4.8 23.40 1.88 23.97Spinal cord 11.1 4.5 13.74 1.37 13.53Brainstem 15.1 12.6 13.61 2.14 13.79

Abbreviations: CAT = computer-assisted target volume delinea-tion; CFS = contouring from scratch; CTV = clinical target volume;ROI = region of interest; SD = standard deviation.

Table 3. Contouring variation among participatingphysicians in the case of the patient with nasopharyngeal cancer

CFS CAT-assistedTemplatecontours

ROI



Volume(cm3)

CTV1 140.82 83.93 101.76 7.90 99.94CTV2 325.41 88.89 297.15 18.32 299.19CTV3 94.91 65.55 91.09 9.88 91.57Left parotid 28.94 6.41 26.31 1.84 25.54Right parotid 28.99 4.23 23.36 0.96 23.73Spinal cord 12.69 5.43 11.55 0.78 12.18Brainstem 33.47 12.58 27.08 2.07 26.03Optic chiasm 1.59 0.68 1.82 0.47 1.56Left optic nerve 0.95 0.63 0.73 0.13 0.87Right optic nerve 0.82 0.51 0.86 0.05 0.86Left orbit 9.08 2.12 8.11 0.59 8.40Right orbit 9.78 1.98 8.18 0.36 8.60

Abbreviations: CAT = computer-assisted target volume delinea-tion; CFS = contouring from scratch; CTV = clinical target volume;ROI = region of interest; SD = standard deviation.


Fig. 2. Improvement in contouring variation using computer-assisted contouring for the case of base-of-tongue carcinoma.The top row of images illustrate the degree of variation among the target volumes that were contoured from scratch by thetesting physicians. The bottom row shows the improvement in consistency achieved by using the deformable image-registration system. The first column of images depicts the coronal aspect, whereas the next two columns, marked withnumbers 1 and 2, show the differences between the two methods of contouring from the axial aspect.

(p = 0.012 and p = 0.001, respectively). In contrast, the var-

iation in the volume of normal tissue contours was more con-

sistent, except in the brainstem, which may be associated

with the inherent limitation of CT images in identifying the

anatomic boundaries of the brainstem. Because the focus of

this study is mainly on variation among physicians, not find-

ing the gold standard for practice that needs to be derived

from rectifying consensus among thought leaders using evi-

dence-based approach, it was expected that if contours that

participating physicians generated from scratch (CFS) are

consistent among themselves, then the VOI (either high or

low in value) will be similar when compared with template

contours. We did not observe this trend but instead the results

of the VOI analysis showed both a significant variation in

both BOT and NPC cases (Table 4 and 5). The variation

was reduced in the geometric distribution of perceived

Table 4. Volume overlap index for the case of the patient with base-of-tongue cancer

Volume overlap index (%) Distance disagreement

CFS CAT-assisted CFS CAT-assisted

ROI Mean Max Min Mean Max Min p value Mm SD mm SD p value

CTV1 65 80 47 93 96 88 <0.001 5.2 2.2 1.0 0.3 0.001CTV2 71 83 56 95 97 93 <0.001 5.9 2.7 0.9 0.3 0.002CTV3 53 72 32 91 94 87 <0.001 8.8 6.7 0.8 0.6 0.014Left parotid 71 84 66 89 96 84 <0.001 2.9 0.9 0.7 0.3 0.002Right parotid 77 80 73 91 97 86 <0.001 1.9 0.5 0.6 0.4 0.002Spinal cord 68 74 53 93 100 76 <0.001 2.0 1.1 0.3 0.3 0.004Brainstem 49 60 11 88 100 77 0.018 5.3 4.1 1.0 0.6 0.036Average 61 78 44 90 97 81

Abbreviations: CAT = computer-assisted target volume delineation; CFS = contouring from scratch; CTV = clinical target volume;Max = maximum; Min = minimum; ROI = region of interest; SD = standard deviation.


Table 5. Volume overlap index for the case of the patient with nasopharyngeal cancer

Volume overlap index (%) Distance disagreement

CFS CAT-assisted CFS CAT-assisted

ROI Mean Max Min Mean Max Min p Value mm SD mm SD p Value

CTV1 61 78 21 90 95 84% <0.001 7.9 10.1 1.0 0.6 0.012CTV2 65 79 59 92 96 84 <0.001 8.3 4.8 1.0 0.7 0.005CTV3 32 74 0 89 95 83 <0.001 14.6 9.8 0.8 0.3 0.005Left parotid 74 84 68 90 98 85 <0.001 2.2 0.8 0.5 0.2 0.001Right parotid 78 86 71 92 97 87 <0.001 2.2 0.8 0.4 0.2 0.001Spinal cord 70 83 59 93 100 86 <0.001 1.7 0.8 0.4 0.2 0.001Brainstem 76 88 65 92 98 85 <0.001 2.4 1.0 0.5 0.3 0.002Optic chiasm 12 58 0 71 94 0 0.028 4.2 2.3 1.1 2.6 0.122Left optic nerve 46 64 10 80 96 64 <0.001 1.8 1.9 0.4 0.3 0.088Right optic nerve 43 60 1 86 93 72 <0.001 1. 2.3 0.2 0.1 0.094Left orbit 85 89 80 93 99 88 <0.001 1.0 0.4 0.4 0.2 0.004Right orbit 80 86 74 93 97 90 <0.001 1.4 0.6 0.3 0.1 0.002Average 57 76 40 88 96 76 0.005

Abbreviations: ROI = region of interest; CFS = contouring from scratch; CAT = computer-assisted target volume delineation;Max = maximum; Min = minimum; SD = standard deviation; CTV = Clinical target volume.

anatomic orientation of CTVs and normal structures when

using CAT approach. Examples of target volume delineation

from scratch and with the assistance of the CAT system are

illustrated in Figs. 2 and 3.

Case of BOT cancerTable 4 summarizes the data on the VOI and distance

agreement in the case of the patient with BOT cancer. The

average VOIs for the CTV1, CTV2, and CTV3 when the

Fig. 3. Improvement in contouring variation using computer-assisted contouring (for the case of nasopharyngeal carci-noma). The top row of images illustrate the degree of variation among the target volumes that were contoured from scratchby the testing physicians. The bottom row shows the improvement in consistency achieved by using the deformable imageregistration system. The first column of images depicts the coronal aspect, whereas the next two columns, marked withnumbers 1 and 2, show the differences between the two methods of contouring from the axial aspect.


physicians CFS were 66%, 72%, and 55%, respectively. The

corresponding average VOIs using CAT were 93%, 95%,

and 91%, respectively. The average VOIs for all structures

were 63% and 90% for CFS and CAT, respectively. The

improvement in the VOI when using CAT was significant

(p < 0.001).

The distance agreement improved from 5.2 � 2.2 mm to

1.0 � 0.3 mm for CTV1 using the CAT system; similar im-

provements were seen for CTV2 and CTV3. The distance

agreement of all regions of interest improved from 4.6 �2.8 mm to 0.8 � 0.5 mm (p = 0.002).

Case of NPCTable 5 shows the data on the VOI and distance agreement

for the second patient, who had NPC. The average VOIs for

CTV1, CTV2, and CTV3 for CFS were 63%, 68%, and 34%,

respectively. The same VOIs using CAT were 90%, 92%, and

89%, respectively. The average VOIs for all structures were

59% and 88.5%forCFS andCAT, respectively. Therewas a sig-

nificant improvement in the VOI when using CAT (p = 0.001).

The distance agreement improved from 7.9 � 10.1 mm to

1.0� 0.6 mm for CTV1 using the CAT system, and a similar

improvement was seen for CTV2 and CTV3. The distance

agreement for all regions of interest improved from 4.3 �3.0 mm to 0.6 � 0.5 mm (p = 0.001).

Time saved in contouring with CATThe data on the time saved in contouring with the CAT

system are shown in Tables 6 and 7. The average time re-

quired for the ‘‘frequent’’ physicians to perform CFS on

the BOT case was 44 min; this improved to 27 min using

the CAT. The corresponding values for the ‘‘frequent’’

physicians for the case of NPC were 56 min and 35 min, re-

spectively. This yielded an average time saved for both cases

of 19.5 min.

For the ‘‘sporadic’’ physicians, the average time for per-

forming CFS on the BOT case was 55 min; this improved

to 28 min using the CAT. The corresponding values for ‘‘spo-

radic’’ physicians for contouring the NPC case were 65 min

and 39 min, respectively. This yielded an average time saved

for both cases of 26.5 min.

We found that among the ‘‘frequent’’ physicians, there

was an average timesaving of 26% and 29% on the BOT

and NPC cases between the two methods of contouring.

Among the ‘‘sporadic’’ physicians using CAT to assist

with contouring; however, the timesaving became more

prominent: the average times saved for that group were

47% for the BOT case and 38% for the NPC case.

DISCUSSION

Our CAT system utilizing deformable image registration

method to morph a set of template contours from one refer-

ence image set to the patient’s CT images improved contour-

ing consistency among physicians with varying HN IMRT

experience and shortened the time spent on the laborious con-

touring process. The implication of this approach in clinical

practice can be significant because the use of IMRT is

increasing rapidly. Mell et al. conducted a survey of 500

American Society for Therapeutic Radiology and Oncology–

registered radiation oncologists regarding their use of IMRT

in 2002 and again in 2004. That investigation showed that the

use of IMRT in clinical practice increased from 32% in 2002

to 67.8% in 2004. In addition, among the physicians not

currently using IMRT, nearly 60% said they would start

using it within the next year, and about 90% said that they

would implement it in the future (18, 19).

Nonetheless, implementation of IMRT faces considerable

bottlenecks in most community practices, especially in small

and solo practices. These bottlenecks exist because physicians

in community practice lack experience and comprehensive

training in cross-sectional anatomy so they can delineate tumor

targets on CT images and that the laborious and repetitive con-

touring process requires more time than physicians have avail-

able. In many community practices, a certain proportion of the

contouring process, mainly of normal tissues, is shared with

support personnel, such as trainees, physician’s assistants, or

certified medical dosimetrists. However, an overall shortage

of the workforce has remained constant over the past few

years. For example, the workforce surveys of 2004 and 2005

conducted by the American Association of Medical Dosimet-

rists showed that 35% to 45% of community practices were

short-handed (20, 21).

Growing adoption of IMRT utilization is also anticipated

in clinical practices outside the United States, and there the

workforce shortage will also affect physicians’ time to per-

form the target volume delineation necessary to meet a certain

standard of quality.

Table 6. Time analysis for case of the patient with base of tongue cancer

Frequent HN IMRT physicians Sporadic HN IMRT physicians

Physicianno.

CFS time(min)

CAT time(min) Ratio

Time saving(min)

Physicianno.

CFS time(min)

CAT time(min) Ratio

Time saving(min)

8 24 29 1.21 �5 2 43 24 0.56 196 38 31 0.82 7 1 65 16 0.25 497 45 29 0.64 16 3 50 37 0.74 135 69 20 0.29 49 4 60 35 0.58 25Average 44 27 0.74 17 Average 55 28 0.53 27

Abbreviations: HN = head and neck; IMRT = intensity-modulated radiation therapy; CFS = contouring from scratch; CAT = computer-assisted target volume delineation.


Table 7. Time analysis for case of the patient with nasopharyngeal cancer

‘‘Frequent’’ HN IMRT physicians ‘‘Sporadic’’ HN IMRT physicians

Physicianno.

CFS time(min)

CAT time(min) Ratio

Time saving(min)

Physicianno.

CFS time(min)

CAT time(min) Ratio

Time saving(min)

8 38 33 0.87 5 2 45 35 0.78 106 33 35 1.06 �2 1 76 39 0.51 377 75 40 0.53 35 3 75 38 0.51 375 79 30 0.38 49 4 65 45 0.69 20Average 56 35 0.71 22 Average 65 39 0.62 26

Abbreviations: CAT = computer-assisted target volume delineation; CFS = contouring from scratch; HN = head and neck; IMRT = intensity-modulated radiation therapy.

Variations in target volume determination and delineationSeveral reports have been published on variations and in-

consistencies in target volume determination and delineation

of tumor sites for which IMRT is being used for radiation

delivery (22–27). In an international study for interobserver

variation on target volume determination and delineation in-

volving 20 institutions in the United States, Europe, and Asia,

Hong et al. (8) found remarkable heterogeneity in HN target

delineation. There was a fivefold variation in CTV volumes

for the ipsilateral neck (range, 35–175 cm3; mean, 120 cm3).

Similar variation was identified for treatment volumes in the

contralateral neck if treated. Wide variation also existed in

the specific nodal stations included within the designated

targets (8). In addition, the Scientific Association of Swiss

Radiation Oncology conducted a study of delineated target

volume variation for HN and prostate cancer in 11 radiation

oncology centers (28). The volumes drawn by the 11 physi-

cians varied widely in that study as well. With regard to

a HN case, the mean CTV was 327.0 cm3 (range, 119.7–

601.6 cm3) (28).

In yet another study, Steenbakkers et al. (29) compared

target volume delineation in 10 cases of NPC among 10

physicians from six institutions. The authors compared the

delineated volumes using a three-dimensional distance agree-

ment method and expressed the results in the form of standard

deviations. The observer variation for all patients for the CTV

was 5.9 mm (1 SD). For the CTV, the anatomic region with

the smallest observer variation was the bone region (SD, 4.4

mm; agreement, 47%), and the largest observer variation

occurred in the caudal region (SD, 7.9 mm; agreement, 7%).

Furthermore, there was a large amount of disagreement in

delineation of the normal nasopharynx (SD, 6.2 mm;

agreement, 10%) (29).

If instructions and guidelines for GTV-to-CTV margins

were known to the physicians during manual contouring why

does variation still exist? This indeed points to the Achilles’

heel of IMRT target volume contouring. Potential reasons for

such significant variations are: a) variable knowledge and

interpretation of ICRU definition of CTV resulting from

differences in training and experience among participating

physicians; b) discrepancy in understanding of microscopic

extension of gross tumor and patterns of nodal spreading;

c) different interpretations of cross sectional anatomy on

computed tomography leading to variation in compassing

the regions intended to be contoured; d) incoherence between

knowledge and practice (drawing) that created dissociation

between knowing and doing. Obviously, the advancements

in imaging and treatment-delivery technology have created

a knowledge gap that creates a bottleneck in the spread of

the use of IMRT. However, inconsistency is rooted in the

key differences in training and practice experience. Target

volume delineation is performed on CT images, but many ra-

diation oncologists in community practice lack this training

and expertise. Although the published guidelines seek to pro-

vide standardization for HN target delineation, in practice the

guidelines can pose challenges for routine use. The HN

IMRT technique is relatively new and is accompanied by

a steep learning curve that prompts centers to refine specific

treatment methods as experience is gained. The guidelines

themselves are complex and require considerable study and

subsequent correlation with the three-dimensional HN anat-

omy of the individual patient under treatment. Further,

most guidelines published to date refer to the less complex

node-negative (N0) manifestation, rather than to the node-

positive (N+) one, for which the variation is more complex

and depends on the number, size, location, and regional infil-

tration of metastatic nodes (30).

Moreover, the clinical manifestation of HN cancer with

gross tumor at the primary site and neck nodes can be complex.

Detailed guidelines and instructions that provide three-dimen-

sional volumetric information pertaining to tumor target and

normal tissue delineation are unlikely to fit in a conventional

print format such as journals or books. In this study, as consis-

tent with prior observations detailed above, we also found

significant variation in delineating CTVs from scratch among

eight physicians. To address this challenge, we opted for

a knowledge-based computer-assisted approach (CAT) that al-

lows the experiences of academic researches to be dissemi-

nated to physicians in community practices. As depicted in

Figures 2 and 3, the variation of the contoured volumes was

significantly reduced when the CAT system was implemented.

This phenomenon was evident in both test cases. In addition,

significant improvement in the VOI occurred when CAT

was used for the two cases (p < 0.001 for both).

Time constraintsThe spectrum of presentation of HN cancer patients is

broad when one takes into account the various T and N


stages. This breadth renders the application of simple ana-

tomic guidelines challenging and entails more time spent

for target volume delineation. Two studies have looked at

the time factor involved in delineation. In the study by

Hong et al. (8), the average time that responders spent con-

touring HN targets was 100 min (range, 60–210 min). Re-

sponders, on average, stated that they typically spent just

over 120 min contouring targets for HN IMRT patients in

their practice (range, 45–270 min) (8). In the second study,

which looked at the effect of IMRT on clinical workloads,

Miles et al. (7) reported the mean target volume outlining

time for clinicians to be 2.3 h (range, 0.7–3.5 h). In our study,

to contour the target volumes within the designated CT image

slices, we found that among the experienced physicians, there

was a average timesaving of 26 and 29% on the BOT and

NPC cases, respectively. The time saved by using CAT

became more evident for less experienced physicians: the

average times saved were 47% for the BOT case and 38%

for the NPC case.

Trend in technology developmentIntending to improve consistency and efficiency in IMRT

target volume determination and delineation, various groups

have developed and investigated the feasibility of use of

semiautomatic target delineation tools. Lu et al. (31) looked

at an automatic recontouring method that combined the tech-

niques of deformable registration and surface construction.

Those investigators validated their methods in the setting

of recontouring for four-dimensional radiotherapy (for lung

cancer) and not for primary contouring (31). Kaus et al.(32) investigated it in the setting of HN cancer, but they de-

lineated only the N0 neck. Butler et al. (33) made use of CT

image fusion with maps containing normal HN anatomy, and

found this technique useful for providing information on

normal anatomical structures. These approaches alike will

certainly gain interest in radiation oncology community since

the cost reimbursement for IMRT is a fixed amount for each

HN case regardless of how much time spent on treatment

planning. Any innovative solutions that can facilitate the

learning process for physicians with varying training back-

ground and unclog the IMRT treatment-planning bottlenecks

will improve the quality of patient care and increase opera-

tional efficiency. In this study, we used a novel and robust

time-efficient software tool (CAT) to aid in contouring clin-

ical target volumes for primary disease (CTV1), N+ neck

(CTV2), and N0 neck (CTV3). We have also used a similar

approach in a wide range of disease manifestations in differ-

ent HN subsites, and the preliminary results were promising.

The CAT system can also be applied beyond the HN ana-

tomic site and the feasibility studies for tumors of other body

sites are underway. We need to caution that although with

CAT and reference templates, the variation was significantly

reduced so was precision improved but not necessarily the ac-

curacy which is beyond the scope of the current study. Any

imperfection in either the computation algorithm or template

contours can lead to a systematic deviation yet little variation.

Nevertheless, if implemented cautiously, this approach may

serve as a platform for education and quality assurance in tar-

get volume delineation when executing IMRT treatment pro-

tocols among different physicians, across varying practices,

and for multi-institutional studies.

CONCLUSION

We found that a computer-assisted contouring system us-

ing a deformable image registration method with template

contours was able to reduce the variation in target volume de-

lineation among physicians with different experiences in HN

IMRT while saving contouring time.

REFERENCES

1. Dawson LA, Anzai Y, Marsh L, et al. Patterns of local-regionalrecurrence following parotid-sparing conformal and segmentalintensity-modulated radiotherapy for head and neck cancer.Int J Radiat Oncol Biol Phys 2000;46:1117–1126.

2. Zelefsky MJ, Fuks Z, Hunt M, et al. High dose radiation deliv-ered by intensity modulated conformal radiotherapy improvesthe outcome of localized prostate cancer. J Urol 2001;166:876–881.

3. Zelefsky MJ, Fuks Z, Hunt M, et al. High-dose intensity mod-ulated radiation therapy for prostate cancer: Early toxicity andbiochemical outcome in 772 patients. Int J Radiat Oncol BiolPhys 2002;53:1111–1116.

4. Chao KS, Ozyigit G, Tran BN, et al. Patterns of failure in patientsreceiving definitive and postoperative IMRT for head-and-neckcancer. Int J Radiat Oncol Biol Phys 2003;55:312–321.

5. Eisbruch A, Ship JA, Dawson LA, et al. Salivary gland sparingand improved target irradiation by conformal and intensity mod-ulated irradiation of head and neck cancer. World J Surg 2003;

27:832–837.6. Chao KS, Ozyigit G, Blanco AI, et al. Intensity-modulated radi-

ation therapy for oropharyngeal carcinoma: Impact of tumorvolume. Int J Radiat Oncol Biol Phys 2004;59:43–50.

7. Miles EA, Clark CH, Urbano MT, et al. The impact of introduc-

ing intensity modulated radiotherapy into routine clinical prac-

tice. Radiother Oncol 2005;77:241–246.8. Hong TS, Chappell RJ, Harari PM. Variations in target delinea-

tion for head and neck IMRT: An international multi-institu-

tional study. Int J Radiat Oncol Biol Phys 2004;60:S157.9. Gregoire V, Coche E, Cosnard G, et al. Selection and delinea-

tion of lymph node target volumes in head and neck conformal

radiotherapy. Proposal for standardizing terminology and

procedure based on the surgical experience. Radiother Oncol2000;56:135–150.

10. Chao KS, Wippold FJ, Ozyigit G, et al. Determination and

delineation of nodal target volumes for head-and-neck cancer

based on patterns of failure in patients receiving definitive and

postoperative IMRT. Int J Radiat Oncol Biol Phys 2002;53:

1174–1184.11. Prins-Braam PM, Raaijmakers CP, Terhaard CH. Location of

cervical lymph node metastases in oropharyngeal and hypo-

pharyngeal carcinoma: Implications for cranial border of elec-

tive nodal target volumes. Int J Radiat Oncol Biol Phys 2004;

58:132–138.


12. Eisbruch A, Marsh LH, Dawson LA, et al. Recurrences nearbase of skull after IMRT for head-and-neck cancer: Implicationsfor target delineation in high neck and for parotid gland sparing.Int J Radiat Oncol Biol Phys 2004;59:28–42.

13. Apisarnthanarax S, Elliott DD, El-Naggar AK, et al. Determin-ing optimal clinical target volume margins in head-and-neckcancer based on microscopic extracapsular extension of meta-static neck nodes. Int J Radiat Oncol Biol Phys 2006;64:678–683.

14. Gregoire V, Eisbruch A, Hamoir M, et al. Proposal for thedelineation of the nodal CTV in the node-positive and the post-operative neck. Radiother Oncol 2006;79:15–20.

15. Thirion JP. Image matching as a diffusion process: An analogywith Maxwell’s demons. Med Image Anal 1998;2:243–260.

16. Wang H, Dong L, Lii MF, et al. Implementation and validationof a three-dimensional deformable registration algorithm fortargeted prostate cancer radiotherapy. Int J Radiat Oncol BiolPhys 2005;61:725–735.

17. Breu H, Gil J, Kirkpatrick D, Werman M. Linear time Euclideandistance transform algorithms. IEEE Transactions on PatternAnalysis and Machine Intelligence 1995;17:529–533.

18. Mell LK, Roeske JC, Mundt AJ. A survey of intensity-modu-lated radiation therapy use in the United States. Cancer 2003;98:204–211.

19. Mell LK, Mehrotra AK, Mundt AJ. Intensity-modulatedradiation therapy use in the U.S., 2004. Cancer 2005;104:1296–1303.

20. Chu RY. AAMD salary and workforce survey. AmericanAssociation of Medical Dosimetrists; 2004.

21. Chu RY. AAMD salary and workforce survey. AmericanAssociation of Medical Dosimetrists; 2005.

22. Logue JP, Sharrock CL, Cowan RA, et al. Clinical variability oftarget volume description in conformal radiotherapy planning.Int J Radiat Oncol Biol Phys 1998;41:929–931.

23. Fiorino C, Reni M, Bolognesi A, et al. Intra- and inter-observervariability in contouring prostate and seminal vesicles: Implica-

tions for conformal treatment planning. Radiother Oncol 1998;47:285–292.

24. Cazzaniga LF, Marinoni MA, Bossi A, et al. Interphysician var-iability in defining the planning target volume in the irradiationof prostate and seminal vesicles. Radiother Oncol 1998;47:293–296.

25. Saarnak AE, Boersma M, van Bunningen BN, et al. Inter-observer variation in delineation of bladder and rectum contoursfor brachytherapy of cervical cancer. Radiother Oncol 2000;56:37–42.

26. Hurkmans CW, Borger JH, Pieters BR, et al. Variability in tar-get volume delineation on CT scans of the breast. Int J RadiatOncol Biol Phys 2001;50:1366–1372.

27. Collier DC, Burnett SS, Amin M, et al. Assessment of consis-tency in contouring of normal-tissue anatomic structures.J Appl Clin Med Phys 2003;4:17–24.

28. Jeanneret-Sozzi W, Moeckli R, Valley JF, et al. The reasons fordiscrepancies in target volume delineation: A SASRO study onhead-and-neck and prostate cancers. Strahlenther Onkol 2006;182:450–457.

29. Steenbakkers RDJ, Fitton I, Deurloo K, et al. Observer variationin delineation of nasopharyngeal carcinoma for radiotherapy.A 3-D analysis. Int J Radiat Oncol Biol Phys 2004;60:S160.

30. Hong TS, Tome WA, Harari PM. Intensity-modulated radiationtherapy in the management of head and neck cancer. Curr OpinOncol 2005;17:231–235.

31. Lu W, Olivera GH, Chen Q, et al. Automatic re-contouring in4D radiotherapy. Phys Med Biol 2006;51:1077–1099.

32. Kaus MR, O’Sullivan B, Mansouri A, et al. Evaluation ofa semi-automated segmentation method for delineation of or-gans at risk and lymph node target volumes in head and neckradiotherapy planning. Int J Radiat Oncol Biol Phys 2006;66:S48.

33. Butler ESV, Yogeswaren ST, Childress C, et al. Novel strate-gies to improve target delineation for head and neck cancers.Int J Radiat Oncol Biol Phys 2004;60:S158.

Documents

Reduce in Variation and Improve Efficiency of Target Volume Delineation by a Computer-Assisted System Using a Deformable Image Registration Approach