Publications

Detailed Information

Leveraging Deep Learning Techniques for Lung Nodule Detection and Segmentation : 딥 러닝 기술을 활용한 폐 결절 탐지 및 분할

DC Field Value Language
dc.contributor.advisor신영길-
dc.contributor.author무하마드-
dc.date.accessioned2023-11-20T04:24:31Z-
dc.date.available2023-11-20T04:24:31Z-
dc.date.issued2023-
dc.identifier.other000000177857-
dc.identifier.urihttps://hdl.handle.net/10371/196500-
dc.identifier.urihttps://dcollection.snu.ac.kr/common/orgView/000000177857ko_KR
dc.description학위논문(박사) -- 서울대학교대학원 : 공과대학 컴퓨터공학부, 2023. 8. 신영길.-
dc.description.abstractThis dissertation offers an exhaustive examination of innovative deep learning methodologies developed to enhance the detection and segmentation of lung nodules in thoracic computed tomography (CT) scans, thus enabling early diagnosis and more efficacious treatment planning for lung cancer. Given that lung cancer is the most fatal cancer worldwide, the early detection and accurate analysis of lung nodules become imperative for interventions that can save lives. However, the heterogeneity of lung nodules and their varied size and shape present substantial challenges for the creation of highly precise computer-aided detection (CADe) systems.

Initially, this dissertation introduces a novel multi-encoder-based self-distilled network. This network uses multimodal convolutional neural networks (CNNs) and merges three types of inputs: sub-volumes of CT scans, along with maximum intensity projection (MIP) images in both forward and backward directions. The approach employs a single network for lung nodule detection and false positive reduction, providing an efficient computational method that outperforms existing single and dual-stage CADe systems.

Following the detection of lung nodules, accurate segmentation is crucial for evaluating the malignancy level of the nodule and subsequent treatment planning. Accordingly, this dissertation presents three distinct frameworks for lung nodule segmentation. All three frameworks independently examine the existence of a nodule in adjacent slices using a 2D region of interest (ROI) as input. First, a unique adaptive ROI algorithm is devised, which employs the 2D segmentation of the nodule in the current slice to estimate the ROI position and size for the surrounding slices. Initially, the A-ROI algorithm is used with Residual UNet architecture with multi-view analysis for nodule segmentation. Later, the dual-encoder-based hard attention network (DEHA-Net) is employed to counteract the degradation in performance caused by the rescaling of the input image. Finally, an efficient end-to-end framework, the multi-encoder-based adaptive hard attention architecture (MESAHA-Net) is proposed. This framework integrates three encoding paths and enables slice-by-slice 2D segmentation of lung nodules. MESAHA-Net has proven its robustness against various lung nodule types and has outperformed prior state-of-the-art techniques in terms of segmentation accuracy and computational complexity, proving its aptness for real-time clinical applications.

The proposed algorithms pave the way for exploring research areas such as refining the lung nodule diagnosis pipeline to include lung nodule classification based on texture and malignancy level. These techniques can also be applied to develop a comprehensive pipeline for the automatic follow-up of lung nodules in subsequent scans for the same patient. Such advancements hold the promise to substantially boost radiologist productivity and patient survival rates.
-
dc.description.abstract이 학위 논문은 흉부 컴퓨터 단층촬영(CT) 스캔에서 폐 결절의 탐지와 분할을 향상시키기 위해 설계된 혁신적인 딥러닝 기법에 대한 포괄적인 탐구를 제시합니다. 이를 통해 폐암, 세계에서 가장 치명적인 암 형태의 조기 진단과 보다 효과적인 치료 계획이 가능해집니다. 그러나 폐 결절의 이종성 및 다양한 크기와 형태는 고도로 정확한 컴퓨터 보조 탐지(CADe) 시스템 개발에 중요한 도전을 제기합니다.

이 학위 논문은 컨볼루션 신경망(CNNs)을 활용하고 CT 스캔의 부분 볼륨과 전방 및 후방 방향의 최대 강도 투영(MIP) 이미지를 통합하는 세 가지 입력 유형을 활용하는 독특한 멀티 인코더 기반 자기 증류 네트워크를 소개합니다. 이 프레임워크는 폐 결절 탐지와 거짓 양성 감소 모두에 대해 단일 네트워크를 사용하여, 기존의 단일 및 이중 단계 CADe 시스템에 대한 우수한 성능을 보여주는 동시에 계산 효율성을 제공합니다.

폐 결절 탐지 후에는 결절의 악성 수준을 평가하고 후속 치료를 계획하기 위해 정확한 분할이 필수적입니다. 이에 따라, 이 학위 논문에서는 폐 결절 분할을 위한 세 가지 독특한 프레임워크를 제안합니다. 세 프레임워크 모두 2D 관심 영역(ROI)을 입력으로 사용하면서 주변 슬라이스에서 결절의 존재를 독립적으로 조사합니다. 첫째, 현재 슬라이스의 결절 2D 분할을 활용하여 주변 슬라이스의 ROI 위치와 크기를 추정하는 새로운 적응 ROI 알고리즘이 고안되었습니다. 초기에는 A-ROI 알고리즘이 멀티 뷰 분석을 사용하여 결절을 분할하는 Residual UNet 아키텍처와 함께 사용되었습니다. 후에, 입력 이미지의 재확대로 인한 성능 저하를 완화하는 데 사용되는 이중 인코더 기반 하드 어텐션 네트워크(DEHA-Net)가 사용되었습니다. 마지막으로, 효율적인 엔드-투-엔드 프레임워크인 멀티 인코더 기반 적응 하드 어텐션 아키텍처(MESAHA-Net)가 제안되었습니다. 이는 세 가지 인코딩 경로를 통합하고 폐 결절의 슬라이스 별 2D 분할을 지원합니다. MESAHA-Net은 다양한 폐 결절 유형에 대한 강건성을 보여주고 분할 정확성과 계산 복잡성 면에서 이전의 최신 기술을 능가하였으며, 따라서 실시간 임상 응용에 적합함을 입증하였습니다.

제안된 알고리즘은 텍스처와 악성 수준에 따른 폐 결절 분류를 포함하여 파이프라인을 개선하고 동일한 환자의 후속 스캔에서 폐 결절의 자동 후속 관리를 위한 완전한 파이프라인을 개발하는 연구 방향을 탐색하는 데 활용될 수 있습니다. 이러한 진전은 방사선 전문가의 생산성과 환자의 생존률을 크게 향상시킬 것으로 약속하며, 동시에 폐 결절 탐지와 분할 분야에 가치 있는 기여를 제공합니다.
-
dc.description.tableofcontentsChapter 1 Introduction 1
1.1 Background and motivation 1
1.2 Problem statement 6
1.3 Main contributions 10
1.4 Contents and organization 14
Chapter 2 Review of Existing Techniques in Lung Nodule Detection and Segmentation 15
2.1 Overview 15
2.2 Convolutional neural networks 16
2.2.1 Architectures of convolutional neural networks 16
2.2.2 Convolutional neural networks for disease detection and segmentation in medical images 29
2.3 Lung nodule detection and segmentation 44
2.3.1 Lung nodule detection systems 46
2.3.2 Lung nodule segmentation systems 59
2.4 Motivation 63
Chapter 3 Development and Assessment of a Multi-Encoder-Based Self-Distilled Network for Lung Nodule Detection 67
3.1 Overview 67
3.2 Bidirectional maximum intensity projection 70
3.3 Multi-encoder-based self-distilled network 75
3.3.1 Dense block 76
3.3.2 Encoder block 77
3.3.3 Decoder block 78
3.4 Self-distillation mechanism 79
3.5 False positive reduction using auxiliary detectors 81
3.6 Experimental setup 82
3.6.1 Datasets. 82
3.6.2 Data preprocessing 83
3.6.3 Training Strategy 84
3.7 Experimental Results 86
3.7.1 Overview 86
3.7.2 Ablation study 91
3.8 Discussion 99
Chapter 4 Designing a Residual UNet for Lung Nodule Segmentation: Integration of Scaled Adaptive Regions of Interest 102
4.1 Overview 102
4.2 Stage I: 2D nodule segmentation on axial view by using A-ROI algorithm 105
4.2.1 Adaptive ROI algorithm 105
4.2.2 Deep residual U-Net architecture 111
4.2.3 Loss function 113
4.3 Stage II: Multi-view analysis of lung nodule using residual UNet 113
4.3.1 Multi-view examination 113
4.3.2 Nodule shape reconstruction. 114
4.4 Experimental setup 115
4.4.1 Dataset and Preprocessing 115
4.4.2 Implementation Details 117
4.4.3 Performance Metrics 117
4.5 Analysis of Results 119
4.5.1 Sensitivity of Performance to the RT Value 119
4.5.2 Overall performance 120
4.5.3 Robustness analysis 121
4.5.4 Visual analysis 122
4.6 Constraints of A-ROI Algorithm 126
4.7 Discussion 127
Chapter 5 Advancing Lung Nodule Segmentation with Dual-Encoder-Based Hard Attention Network 129
5.1 Overview 129
5.2 Adaptive ROI mechanism and dual-encoder-based hard attention network 134
5.2.1 Dual-encoder-based hard attention network 134
5.2.2 Adaptive ROI algorithm 136
5.2.3 Ensembling mechanism 137
5.3 Experimental setup and implementation details 138
5.3.1 Loss function 138
5.3.2 Dataset 139
5.3.3 Data pre-processing 139
5.3.4 Details of Implementation and Training Approach 141
5.3.5 Assessment Metrics 141
5.4 Experimental results 142
5.4.1 Overall performance analysis 142
5.4.2 Robustness analysis 143
5.4.3 Qualitative analysis 145
5.5 Discussion 147
Chapter 6 Fine-Tuning Segmentation: Implementation of a Multi-Encoders-Based Self-Adaptive Hard Attention Network 149
6.1 Overview 149
6.2 MESAHA-Net: Multi-encoder-based self-adaptive hard attention network 151
6.2.1 Encoder block 152
6.2.2 Self-adaptive hard attention mechanism 153
6.2.3 Attention block 155
6.2.4 Decoder block 156
6.2.5 Loss function 157
6.3 Experimental Setup 159
6.3.1 Dataset and preprocessing 159
6.3.2 Training and inference strategy 163
6.3.3 Parameters for Evaluation 164
6.4 Results and Discussion 166
6.4.1 Overall performance 166
6.4.2 Robustness analysis 171
6.4.3 Computational time analysis 173
6.4.4 Qualitative analysis 176
6.5 Discussion 178
Chapter 7 A Comparative Discussion on Developed Detection and Segmentation Techniques 180
7.1 Computer-aided lung nodule detection system 181
7.2 Computer-aided lung nodule segmentation systems 184
Chapter 8 Conclusion and Directions for Future Research 191

초록 221
-
dc.format.extentxxii, 222-
dc.language.isoeng-
dc.publisher서울대학교 대학원-
dc.subjectLung nodule detection-
dc.subjectlung nodule segmentation-
dc.subjectself-distillation-
dc.subjectmaximum intensity projection-
dc.subjectadaptive ROI algorithm-
dc.subject.ddc621.39-
dc.titleLeveraging Deep Learning Techniques for Lung Nodule Detection and Segmentation-
dc.title.alternative딥 러닝 기술을 활용한 폐 결절 탐지 및 분할-
dc.typeThesis-
dc.typeDissertation-
dc.contributor.AlternativeAuthor무하마드 우스만-
dc.contributor.department공과대학 컴퓨터공학부-
dc.description.degree박사-
dc.date.awarded2023-08-
dc.identifier.uciI804:11032-000000177857-
dc.identifier.holdings000000000050▲000000000058▲000000177857▲-
Appears in Collections:
Files in This Item:

Altmetrics

Item View & Download Count

  • mendeley

Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.

Share