Open Li Xingyan Thoracic.Pdf
Total Page:16
File Type:pdf, Size:1020Kb
THE PENNSYLVANIA STATE UNIVERSITY SCHREYER HONORS COLLEGE SCHOOL OF ELECTRICAL ENGINEERING AND COMPUTER SCIENCE THORACIC BONE SEGMENTATION FROM 3D CHEST CT SCANS XINGYAN LI SPRING 2018 A thesis submitted in partial fulfillment of the requirements for a baccalaureate degree in Electrical Engineering with honors in Electrical Engineering Reviewed and approved∗ by the following: William E. Higgins Distinguished Professor of Electrical Engineering Thesis Supervisor Julio Urbina Associate Professor of Electrical Engineering Honors Advisor ∗Signatures are on file in the Schreyer Honors College Abstract Segmentation of bone structures, such as ribs, spine, and sternum, in chest CT scans is important for anatomical analysis. Segmented bone structures can serve as the reference for locating organs in the thoracic cavity, such as lymph node stations. In this thesis, we aim to develop a method for segmenting ribs, spine, and sternum from chest CT images. The method can be decomposed into seven major steps. The method was evaluated using a 10-case CT scan dataset. 2 cases of CT scans were used in the parameter-sensitivity test, and the other 8 cases were used for testing the optimized method. The method achieves Dice and Jaccard similarity values greater than 80 %, and Coverage similarity values greater than 90 %. i Table of Contents Chapter 1 Introduction1 1.1 Background...............................1 1.2 Problem Statement...........................2 1.3 Overview of the Paper.........................3 Chapter 2 Methods4 2.1 Image Thresholding...........................8 2.2 Spinal Canal Centerline........................9 2.2.1 Seed Point of Spinal Canal Centerline.............9 2.2.2 2D Distance Transform for Spinal Canal Tracing....... 11 2.3 Rib Seed Region Identification..................... 14 2.4 Rib Segmentation............................ 20 2.5 Spine Segmentation........................... 23 2.6 Sternum Segmentation......................... 23 2.6.1 Lung Segmentation....................... 24 2.6.2 Sternum Segmentation..................... 25 2.7 Post-Processing............................. 28 2.8 Bone Structures Combined....................... 28 Chapter 3 Results 29 3.1 Test Data and Similarity Measures.................. 29 3.2 Test on Case 21405.3.......................... 31 3.3 Test on Case 20349.3.33........................ 37 3.4 Test on Other Cases.......................... 44 3.4.1 Case 1: 21405.13........................ 44 3.4.2 Case 2: 21405.64........................ 47 3.4.3 Case 3: 21405.66........................ 50 3.4.4 Case 4: 20349.3.7........................ 52 3.4.5 Case 5: 20349.3.15....................... 52 3.4.6 Case 6: 20349.3.22....................... 53 ii 3.4.7 Case 7: 20349.3.29....................... 54 3.4.8 Case 8: 20349.3.35....................... 54 Chapter 4 Conclusion 56 Bibliography 57 iii List of Tables 2.1 Summary of CT scan database.....................5 2.2 Summary of Parameters.........................5 2.3 CT scans used for determining parameter default value.......6 3.1 CT scan dataset............................. 30 3.2 Summary of CT scan database..................... 30 3.3 Similarity results between Gribspine and the segmentation of rib and spine. Case 21405.3. Trib = 250. Tsternum = 180............ 31 3.4 Similarity results between Gsternum and sternum segmentation. Trib = 250, Tspine = 220. Case 21405.3................... 34 3.5 Similarity results between Gribspine and segmentation of rib and spine. Tsternum = 200, Tspine = 220. Case 21405.3.......... 36 3.6 Good value ranges of Tspine, Tsternum, Trib. Case 21405.3....... 36 3.7 Similarity results between Gribspine and the segmentation of rib and spine. Trib = 250, Tsternum = 180. Case 20349.3.33.......... 37 3.8 Similarity results between Gsternum and sternum segmentation. Trib = 250, Tspine = 210. Case 20349.3.33................. 39 3.9 Similarity results between Gribspine and segmentation of rib and spine. Tsternum = 170, Tspine = 210. Case 20349.3.33......... 42 3.10 Good value range of each parameter. Case 20349.3.33........ 42 iv 3.11 Optimized parameter values. "Parameter" denotes name of the parameter. "Optimal Value" gives the parameter's optimal value.. 44 v List of Figures 2.1 Block diagram of proposed 3D bone segmentation method.....7 2.2 Iunion25, the union of 50 axial sections in Rspican, 657≤z≤706.... 10 2.3 Steps of finding the lowest spinal canal center point......... 11 2.4 Example of the 2D Distance Transform Operation.......... 12 2.5 Five candidates of spinal canal center point in axial section I(·; ·; c−1) 13 2.6 The complete spinal canal centerline................. 14 2.7 Image regions to be considered as rib seed regions.......... 15 2.8 3D image of original Rrawseed ...................... 16 2.9 2D images consists of voxels at a distance of δ from the spinal canal 17 2.10 Identified 24 rib seed regions, Rribseed.................. 20 2.11 Rrib with spine region removed.................... 21 2.12 Rib cage segmentation result...................... 22 spine 2.13 Spine segmentation result R from Rspine ............. 23 lung 2.14 Lung segmentation result R from raw mask Rlung......... 24 2.15 Rsternum with a width of 140 mm ................... 25 T 2.16 Rsternum with 3D connected components posterior to p removed.. 26 vi 2.17 Rsternum with only 3D connected components anterior to the trans- lated Rlung ................................ 27 2.18 Sternum segmentation result Rsternum ................. 27 2.19 Bone structures combined....................... 28 3.1 Parameter-sensitivity test results of Tspine on case 21405.3...... 32 3.2 Segmentation of rib and spine. Tspine = 200 & Tspine = 210. Case 21405.3.................................. 33 3.3 Parameter-sensitivity test results of Tsternum on case 21405.3.... 35 3.4 Parameter-sensitivity test results of Trib on case 21405.3....... 35 3.5 Parameter-sensitivity test results of Tspine on case 20349.3.33.... 38 3.6 Parameter-sensitivity test results of Tsternum on case 20349.3.33... 40 3.7 Segmentation of sternum. Tsternum = 180 & Tsternum = 190. Case 20349.3.33................................ 41 3.8 Parameter-sensitivity test results of Trib on case 20349.3.33..... 43 3.9 Segmentation of rib with scapula. Trib = 200, Trib = 260. Case 20349.3.33................................ 43 3.10 Segmented rib cage. Case 21405.13.................. 45 3.11 Spine segmentation result. Case 21405.13............... 45 3.12 Sternum segmentation result. Case 21405.13............. 46 3.13 Complete bone structure segmentation. Case 21405.13....... 46 3.14 Segmented rib cage. Case 21405.64.................. 47 3.15 Spine segmentation result. Case 21405.64............... 48 3.16 Sternum segmentation result. Case 21405.64............. 48 vii 3.17 Complete bone structure segmentation. Case 21405.64....... 49 3.18 Complete bone structure segmentation. Case 21405.66....... 50 3.19 Segmented rib cage with a missing rib. Case 21405.66........ 51 3.20 Spine segmentation result. Case 21405.66............... 51 3.21 Complete bone structure, Rcomplete. Case 20349.3.7......... 52 3.22 Segmetation of complete bone structure, Rcomplete. Case 20349.3.15 53 3.23 Segmetation of complete bone structure, Rcomplete. Case 20349.3.22 53 3.24 Segmetation of complete bone structure, Rcomplete. Case 20349.3.29 54 3.25 Segmetation of complete bone structure, Rcomplete. Case 20349.3.35 54 viii Chapter 1 | Introduction 1.1 Background Lung cancer has been the leading cause of cancer death in both men and women in the United States, accounting for 27 % of cancer deaths in the 2014 [1]. Accu- rate diagnosis and staging of lung cancer is crucial, therefore, in determining the right treatment course [2]. During the past few decades, graduate students and researchers in Multidimensional Image Processing Lab (MIPL) have done work driven by problems in high-resolution 3D medical image processing, such as lung cancer staging. The lab's current focus is on virtual endoscopy and image-guided endoscopy. To be more specific, for example, bronchoscopy is a major step in lung cancer staging. However, bronchoscopy without a guidance system would require physicians to have good skills in performing it. A lot of work has been done in the lab to realize an advanced guidance system for bronchoscopy, which can sig- nificantly improve the efficiency of bronchoscopy procedures. Another problem that has been addressed is the thoracic cavity definition for 3D PET (Positron Emission Tomography) / CT (X-Ray computed tomography) analysis and visual- ization [3]. Normally, thoracic cavity is defined as central chest region above the diaphragm, and it is enclosed by the rib cage, spine and sternum. In fact, a good understanding of bone structures, such as ribs, spine, and sternum, would make a better understanding of thoracic cavity in terms of image segmentation. Bone structures such as ribs can be used as the reference for locating organs and other features within the chest [4]. Bone segmentations from CT scans have not been paid much attention in MIPL at Penn State. Cheirsilp et al. presented a method that mainly utilizes topological 1 operations and several known landmarks to segment ribs, spine, and sternum [3]. However, a lot of work has been done by researchers outside MIPL. Zhang et al. proposed a recursive tracking process of identifying rib components in every 2D coronal section [5]. Their method requires a lot of computations when it comes to a CT scan with more than five hundred coronal sections. Klinder et al. proposed a method that