Enhancing Lung Cancer Detection : Leveraging CT Features for Identifying Nodules of Various Sizes in Chest X-Ray Images /Samar Ibrahim Antar Awwad Ramadan

Supervisor:
Prof. Mustafa Elattar

Thesis (M.A.)—Nile University, Egypt, 2024 .

"Includes bibliographical references"

Contents:
Table of Contents
ACKNOWLEDGEMENT................................................................................................... i
ABSTRACT........................................................................................................................ ii
Table of Contents............................................................................................................... iii
List of Figures................................................................................................................... vii
List of Tables .......................................................................................................................x
List of Equations................................................................................................................ xi
List of Abbreviation......................................................................................................... xiii
Chapter 1 INTRODUCTION...............................................................................................1
1.1. MOTIVATION....................................................................................................1
1.2. PROBLEM DEFINITION...................................................................................3
1.3. CONTRIBUTION................................................................................................4
1.4. THESIS OUTLINES ...........................................................................................5
Chapter 2 BACKGROUND.................................................................................................7
2.1. OVERVIEW ........................................................................................................7
2.2. CLINICAL BACKGROUND..............................................................................8
2.2.1. Respiratory System..........................................................................................8
2.2.2. Lung Cancer...................................................................................................10
2.2.3. Nodule Definitions and Different size of nodules .........................................10
2.2.4. Types of Nodules (Benign vs malignant) ......................................................13
2.2.5. Statistics of Lung Cancer Spread...................................................................15
2.3. LUNG CANCER DIAGNOSIS.........................................................................16
2.4. SUMMARY.......................................................................................................18
Chapter 3 RELATED WORK ...........................................................................................19
3.1. LUNG CANCER DETECTION & SEGMENTATION IN CT........................19
3.2. LUNG CANCER DETECTION & SEGMENTATION IN X-ray....................25
Chapter 4 DATA AND DATASET CREATION..............................................................31
4.1. PUBLIC DATASETS........................................................................................31
4.1.1. Chexpert.........................................................................................................31
4.1.2. The Japanese Society of Radiological Technology .......................................31
4.1.3. Montgomery (NLM-MC)...............................................................................32
4.1.4. Covid-19 and Lung Opacity Datasets............................................................32
iv
4.1.5. LIDC-IDRI.....................................................................................................33
4.1.6. Nodules Chest X-rays (LIDC-IDRI)..............................................................34
4.2. SIMULATED DATASET .................................................................................34
4.2.1. Simulated Data for Image Enhancement .......................................................34
4.3. CREATE SYNTHESIZED 2D CXR FROM 3D THORAX CT.......................35
4.4. ANNOTATION .................................................................................................38
4.5. CORRECTION FOR BOUNDING BOX ANNOTATION AFTER
AUGMENTATION........................................................................................................38
4.6. SUMMARY.......................................................................................................40
Chapter 5 PROPOSED FRAMEWORK ...........................................................................42
5.1. OVERVIEW ......................................................................................................42
5.2. IMAGE PRE-PROCESSING ............................................................................42
5.2.1. Convert CXR images to grayscale and Crop to square..................................43
5.2.2. Adaptive Histogram Equalization..................................................................44
5.2.3. Augmentation.................................................................................................45
5.2.4. Expanding Montgomery Dataset by Merging Lungs.....................................45
5.2.5. The Laplacian of Gaussian (LoG)..................................................................46
5.2.6. Gamma Correction.........................................................................................46
5.2.7. CT Pre-processing..........................................................................................47
5.3. IMAGE ENHANCEMENT...............................................................................49
5.3.1. Proposed image enhancement approach VAE + U-Net:................................49
5.4. LUNG SEGMENTATION................................................................................51
5.5. BONE SUPPRESSION .....................................................................................53
5.5.1. Segmenting Bones from DRRs Using U-Net ................................................53
5.5.2. Generating Rib and Difference images for JSRT ..........................................55
5.5.3. Proposed Latent Boosted Pix2Pix++ Architecture (LB-Pix2Pix++).............56
5.6. CLASSIFICATION ...........................................................................................60
5.7. EXPLAINABLE AI (XAI)................................................................................61
5.7.1. Class Activation Mapping (CAM).................................................................63
5.7.2. Gradient-Weighted Class Activation Mapping (Grad-CAM)........................64
5.8. LOCALIZATION..............................................................................................67
Chapter 6 EXPERIMENTAL SETTINGS ........................................................................68
v
6.1. IMAGE ENHANCEMENT...............................................................................68
6.1.1. Overview........................................................................................................68
6.1.2. VAE Objective Function Optimization and Training....................................68
6.2. LUNG SEGMENTATION................................................................................69
6.2.1. Overview........................................................................................................69
6.2.2. Loss Functions...............................................................................................70
6.2.3. Evaluation matrices........................................................................................71
6.3. BONE SUPPRESSION .....................................................................................72
6.3.1. Bone Segmentation Training .........................................................................72
6.3.2. LB-Pix2Pix++ for Bone Suppression ............................................................72
6.3.3. Evaluation Protocol........................................................................................74
6.4. CLASIFICATION .............................................................................................75
6.4.1. Without data augmentation............................................................................75
6.4.2. With applying data augmentation ..................................................................76
6.4.3. Generalization on JSRT data .........................................................................77
6.5. LOCALIZATION..............................................................................................77
6.5.1. Bounding box estimation (detection).............................................................77
6.5.2. The segmentation model (2nd technique)......................................................78
Chapter 7 RESULTS..........................................................................................................80
7.1. IMAGE ENHANEEMENT ...............................................................................80
7.1.1. Simulation of extra/ halting structures data ...................................................80
7.1.2. The training outcome .....................................................................................80
7.2. LUNG SEGMENTATION................................................................................82
7.3. TESTING OF VAEU-NET AND RESU-NET++.............................................84
7.4. BONE SUPPRESSION .....................................................................................85
7.4.1. Ablation Study ...............................................................................................85
7.4.2. Comparison to the State-of-the-Art ...............................................................89
7.5. CLASSIFICATION ...........................................................................................90
7.5.1. Without data augmentation............................................................................90
7.5.2. With data augmentation .................................................................................90
7.5.3. Generalization on JSRT data & Performance Comparison to Literature ......91
7.5.4. Grad-CAM for the JSRT dataset....................................................................92
vi
7.5.5. Testing on unseen dataset ..............................................................................95
7.6. LOCALIZATION..............................................................................................96
7.6.1. Bounding box estimation (Detection) on JSRT.............................................96
7.6.2. Testing Bounding box detection model on unseen datasets ..........................97
7.6.3. The segmentation model (2nd technique) on JSRT.......................................99
7.6.4. Testing the segmentation model on DRRs...................................................100
Chapter 8 DISCUSSION .................................................................................................103
8.1. LUNG SEGMENTATION..............................................................................103
8.2. TESTING OF VAEU-NET AND RESU-NET++...........................................104
8.3. BONE SUPPRESSION ...................................................................................104
8.4. CLASSIFICATION .........................................................................................107
8.5. LOCALIZATION............................................................................................108
Chapter 9 CONCLUSION AND FUTURE WORK........................................................110
9.1. CONCLUSION................................................................................................110
9.2. FUTURE WORK.............................................................................................111
REFERENCES ................................................................................................................112
Appendix A EXPLAINABILITY....................................................................................121
A.1. CLASS-ENHANCED ATTENTIVE RESPONSE (CLEAR)............................121
A.2. DECONVOLUTIONAL NETWORKS .............................................................121
A.3. LAYER-WISE RELEVANCE PROPAGATION (LRP)...................................123
A.4. DIFFERENCE OF LRP FROM GRAD-CAM AND CAM...............................124
A.5. LRP RULES FOR DEEP RECTIFIER NETWORKS .......................................124
A.6. DEEP LEARNING IMPORTANT FEATURES (DEEPLIFT) .........................125

Abstract:
The high-resolution capability of thoracic Computerized Tomography (CT)
scans enables detailed visualization of lung structures, making it possible to
detect even small abnormalities. This level of detail is especially crucial for
identifying early-stage lung nodules, which may not be visible or
distinguishable in other imaging modalities. However, CT scans have
limitations like cost, access, radiation exposure, and lack of portability. Chest
X-rays (CXRs) are 2D projective images where objects may not appear clear
and can overlap, leading to limited visibility of lung nodules. Deep learning
algorithms have the potential to address these limitations. This thesis proposes
leveraging CT volume features to improve the detection and localization of
lung cancer in CXR, thus enabling radiologists to diagnose lung cancer using
Chest X-ray images more confidently. This could be done by finding CT
features that define small, medium, and large-sized nodules and projecting
those features onto CXR data. The proposed detection system includes all
phases and subtasks required for CXR-based lung cancer identification. Due
to a scarcity of CXR images that contain different nodule sizes, feasibility of
small nodules detection in CXR images is demonstrated by fine-tuning our
predictive model using Digitally Reconstructed Radiographs (DRRs), created
from CT projections. Our proposed method outperforms previous state-of-theart methods, in terms of Mean Intersection Over Union (mIOU) and AUC for
lung cancer detection. Furthermore, the proposed method incorporates
Explainable AI (XAI) module to provide physicians with trustworthy
decision-making tools. It achieves AUC of 0.9762 and MIOU of 0.9025 for
detecting and segmenting nodules on the JSRT dataset, which contains
different sizes of nodules. Additionally, it achieves MIOU of 0.6579 and
0.7646 for small and medium size nodules, respectively.
Keywords:
Chest X-ray, Lung Cancer Segmentation, Computerized Tomography, Different Sizes of
Nodules, Bone Suppression, Chest X-ray Enhancement and DRRs Generation.

Text in English, abstracts in English and Arabic