Detection of Pulmonary Nodules from CT dolejm1/publication/ of Pulmonary Nodules from CT Scans ... indicate a lung cancer and their detection in the early stage improves the ... 2

  • View
    216

  • Download
    2

Embed Size (px)

Text of Detection of Pulmonary Nodules from CT dolejm1/publication/ of Pulmonary Nodules from CT Scans ......

  • CENTER FOR

    MACHINE PERCEPTION

    CZECH TECHNICAL

    UNIVERSITY

    PhD

    TH

    ESIS

    ISSN

    1213-2

    365

    Detection of Pulmonary Nodules

    from CT Scans

    Martin Dolejs

    dolejm1@fel.cvut.cz

    CTUCMP200705

    January 19, 2007

    Available athttp://cmp.felk.cvut.cz/dolejm1/noduledetection/

    Thesis Advisor: Dr. Ing. Jan Kybic

    The CT data were provided by a Faculty Hospital, Motol,Prague.

    Program Scan View was provided and supported by RNDr.Jan Krasensky.

    This work was supported by the Czech Ministry of Healthunder project NR8314-3/2005, and by the Grant Agency ofthe Czech Academy of Sciences under Project 1ET101050403.

    Research Reports of CMP, Czech Technical University in Prague, No. 5, 2007

    Published by

    Center for Machine Perception, Department of CyberneticsFaculty of Electrical Engineering, Czech Technical University

    Technicka 2, 166 27 Prague 6, Czech Republicfax +4202 2435 7385, phone +4202 2435 7637, www: http://cmp.felk.cvut.cz

  • 1

  • 2

  • Abstract

    We present a computer-aided diagnosis (CAD) system to detect small-size(from 2mm to around 10mm) pulmonary nodules from helical CT scans. Apulmonary nodule is a small, round (parenchymal nodule) or worm (juxta-pleural nodule) shaped lesion in the lungs. Both have greater radio-densitythan lung parenchyma, so they appear white on images. Lung nodules mightindicate a lung cancer and their detection in the early stage improves thesurvival rate of patients. CT is considered to be the most accurate imagingmodality for nodule detection. However, the large amount of data per exam-ination makes the interpretation difficult. This leads to omission of nodulesby human radiologist. The presented CAD system is designed to help lowerthe number of omissions and to decrease the time needed to examine thescan by a radiologist. Our system uses two different schemes to locate jux-tapleural nodules and parenchymal nodules respectively. For juxtapleuralnodules, morphological closing and thresholding is used to find nodule can-didates. To locate non-pleural nodule candidates, we use a 3D blob detectorbased on multiscale filtration. To define which of the nodule candidates arein fact nodules, an additional classification step is applied. Linear and multi-threshold classifiers are used. Ellipsoid model is fitted on nodules to providegeometrical features. System was tested on 18 cases (4853 slices) with to-tal sensitivity of 96%, with about 12 false positives/slice. The classificationstep reduces the number of false positives to 9 per slice without significantlydecreasing sensitivity (89.6%). The algorithm was implemented in Matlaband tested under Windows and Unix system. For easy control simple graphicuser interface is included.

    3

  • Abstrakt

    Tato prace se venuje vyvoji programu pro detekci plicnch nodulu. Plicnnoduly jsou kulovite nebo cervovite utvary v plicch, ktere jsou na obrazcchz CT zobrazeny svetlejsmi odstny oproti plicnmu parenchymu. Detekceplicnch nodulu je dulezita, protoze se muze jednat o rakovinna loziska,pritom vcasna diagnoza zasadne ovlivnuje uspesnost lecby vsech druhu zhoub-nych onemocnen. Nas CAD (Computer Aided Diagnostic) system je navrzentak, aby snzil zatzen radiologa pri prohlzen velkeho mnozstv dat, a zarovensnzil pocet prehlednutych nodulu. Program detekuje oddelene kazdy ze dvoudruhu nodulu. Kulovite (parenchymaln) pomoc 3D blobdetektoru (filtracnalgoritmus pracujc ve vce mertkach) a cervovite (juxtapleuraln) pomocprahovan a metod matematicke morfologie. Ke vsem detekovanym objektumpriradme elipsoid popisujc jejich tvar a spoctame jasove statistiky uv-nitr elipsoidu. Pomoc takovehoto popisu klasifikujeme oblasti do dvou trd,jako noduly, nebo jako jine objekty. Pro klasifikaci pouzvame jeden linearna jeden nelinearn (prahovac) klasifikator. Program jsme testovali na 18vysetrench (4853 rezech). Vysledna citlivost samostatnych detektoru byla96%, pri 12 falesne pozitivnch detekcch/rez. Po klasifikaci klesla citlivost na89,6% a pocet falesne pozitivnch na 9/rez. Program byl vyvinut v Matlabua testovan na platforme Windows i Unix.

    4

  • Prohlasen

    Prohlasuji, ze jsem svou diplomovou praci vypracoval samostatne a pouziljsem pouze podklady (literaturu, projekty, SW atd.) uvedene v prilozenemseznamu.

    V Praze dne 19.1.2007,

    5

  • List of Abbreviations

    ANN Artifficial Neural NetworkBW Black and White, binary imageCT Computed TomographyDICOM Digital Imaging and Communication in MedicineF.e. For exampleFLD Fisher Linear DiscriminantFN False NegativeFP False PositiveGUI Graphic User InterfaceHU Hounsfield UnitLoG Laplacian of GaussianSTP Standard Pressure and TemperatureTN True NegativeTP True Positive

    6

  • Contents

    1 Introduction 91.1 Lung Cancer . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91.2 Nodules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101.3 Tomography . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

    1.3.1 Radon Transform . . . . . . . . . . . . . . . . . . . . . 111.3.2 Hounsfield Units . . . . . . . . . . . . . . . . . . . . . 121.3.3 CT Machines . . . . . . . . . . . . . . . . . . . . . . . 13

    2 Nodule Detection 152.1 Problem specification . . . . . . . . . . . . . . . . . . . . . . . 152.2 Previous Work . . . . . . . . . . . . . . . . . . . . . . . . . . 15

    2.2.1 Detection from 2D Data . . . . . . . . . . . . . . . . . 152.2.2 Detection from 3D Data . . . . . . . . . . . . . . . . . 16

    3 Algorithm 193.1 Input Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203.2 Lung Segmentation . . . . . . . . . . . . . . . . . . . . . . . . 203.3 Nodule Candidates Finding . . . . . . . . . . . . . . . . . . . 20

    3.3.1 Parenchymal Nodules . . . . . . . . . . . . . . . . . . . 213.3.2 Juxtapleural Nodules . . . . . . . . . . . . . . . . . . . 213.3.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . 23

    3.4 Nodule Candidates Clasification . . . . . . . . . . . . . . . . . 233.4.1 Problem Description . . . . . . . . . . . . . . . . . . . 243.4.2 The Geometrical Model . . . . . . . . . . . . . . . . . 243.4.3 Model Fitting . . . . . . . . . . . . . . . . . . . . . . . 243.4.4 Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . 27

    4 Methods 294.1 Thresholding . . . . . . . . . . . . . . . . . . . . . . . . . . . 294.2 BW Morphology . . . . . . . . . . . . . . . . . . . . . . . . . 294.3 3D Image Filtering . . . . . . . . . . . . . . . . . . . . . . . . 30

    7

  • 4.3.1 Laplacian of Gaussian . . . . . . . . . . . . . . . . . . 304.4 Scale Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

    4.4.1 3D Blob Detector . . . . . . . . . . . . . . . . . . . . . 314.5 Iterative Maximization Method . . . . . . . . . . . . . . . . . 324.6 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

    4.6.1 Fisher Linear Discriminant . . . . . . . . . . . . . . . . 334.6.2 Multiple Thresholding . . . . . . . . . . . . . . . . . . 34

    5 Implementation 355.1 Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . 355.2 Archive Contents . . . . . . . . . . . . . . . . . . . . . . . . . 355.3 Control GUI . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

    5.3.1 Detection Part of the Dialog . . . . . . . . . . . . . . . 365.3.2 Learning Part of the Dialog . . . . . . . . . . . . . . . 37

    5.4 Matlab Command Line Interface . . . . . . . . . . . . . . . . 385.4.1 Detection . . . . . . . . . . . . . . . . . . . . . . . . . 385.4.2 Learning . . . . . . . . . . . . . . . . . . . . . . . . . . 38

    5.5 Ground Truth Information Reading . . . . . . . . . . . . . . . 395.5.1 Drawings Format . . . . . . . . . . . . . . . . . . . . . 39

    6 Experiments 416.1 Test Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416.2 Evaluation Criteria for Detector Performance . . . . . . . . . . 416.3 Nodule Candidates Detection . . . . . . . . . . . . . . . . . . 426.4 Classification Performance . . . . . . . . . . . . . . . . . . . . 42

    7 Conclusions 477.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

    Image Appendix 49

    Bibliography 59

    8

  • 1 Introduction

    1.1 Lung Cancer

    Lung cancer is one of the leading causes of death in USA [11] and Europe.Surgery, radiation therapy, and chemotherapy are used in the treatment oflung carcinoma. In spite of that, the five-year survival rate for all stagescombined is only 14%. However, early detection helps significantlyit isreported [12] that the survival rate for early-stage localized cancer (stage I)is 49%.

    CT is considered to be the most accurate imaging modality available forearly detection and diagnosis of lung cancer. It allows detecting pathologicaldeposits as small as 1mm in diameter. These deposits are called lung nodules.

    However, the large amount of data per examination makes the interpreta-tion tedious and difficult, leading to a high false-negative rate for detectingsmall nodules. Suboptimal acquisition parameters (e.g. pitch) further de-crease the detection rate. A simulation study demonstrated [1] the overalldetection rate to be only 63% for nodules of 17 mm in diameter. As thesize of the nodule decreased, the sensitivity fell to 48% for nodules smallerthan 3mm, and only 1% of nodules smaller than 1.5 mm in diameter weredetected. Retrospective analysis of CT scans often shows undetected noduleson the initial scans of oncological patients [2].

    Image processing and visualization techniques for volumetric CT datasets may improve the radiologists ability to detect small lung nodules. Forexample, reconstruction of CT images with narrow interscan spacing [3] andinterpretation of images using cine rather than film-b

Recommended

View more >