Publications

Detailed Information

Optimization and Learning for Graph-based Blood Vessel Segmentation : 혈관 영역화를 위한 그래프 기반의 최적화 및 학습 기법 연구

DC Field Value Language
dc.contributor.advisor이경무-
dc.contributor.author신승연-
dc.date.accessioned2019-10-21T02:24:32Z-
dc.date.available2019-10-21T02:24:32Z-
dc.date.issued2019-08-
dc.identifier.other000000157946-
dc.identifier.urihttps://hdl.handle.net/10371/162016-
dc.identifier.urihttp://dcollection.snu.ac.kr/common/orgView/000000157946ko_KR
dc.description학위논문(박사)--서울대학교 대학원 :공과대학 전기·컴퓨터공학부,2019. 8. 이경무.-
dc.description.abstractAnalyzing blood vessels in medical images is crucial for diagnosis and treatment of diverse diseases. Blood vessel segmentation is a primary task for the quantitative analysis, and there have been many attempts to develop an automatic vessel segmentation method. However, the quality is still not guaranteed in terms of clinical application.
The goal of this dissertation is to propose new vessel segmentation methods for diverse problem settings, which include different input modalities, e.g., image and video, and different desired outputs, e.g., binary segmentation and detailed vessel type classification. In particular, with a hypothesis that the vessel network structure is better handled in graph space than in image space due to its nature, all proposed methods are developed based on respective graphical representation of vessel structures.
First, a method to compute vessel correspondences between frames is proposed for video segmentation. The optimal correspondences are acquired by enforcing local appearance similarity and structural consistency on a constructed vessel structure graph. Compared to solving a dense correspondence problem directly on images, it not only reduces the search space but also overcomes ambiguities by the aperture problem. Second, a novel vessel segmentation method based on a graph neural network is proposed, where convolutions are defined on a vessel structure graph for more direct and longer interaction between vessel neighborhoods. Finally, a topology-aware vessel segmentation method is also proposed. Topologically consistent vessel segmentation is achieved by exploiting the topology estimated using a novel deep learning network. Comparative evaluations are performed on multiple retinal image datasets and a coronary artery X-ray angiography dataset, showing the competitiveness of the proposed methods.
-
dc.description.abstract의료 영상 내의 혈관 분석은 다양한 질병을 진단하고 치료하는데 있어 매우 중요한 의미를 가진다. 혈관 영역화는 혈관의 정량적 분석을 위한 과정으로써 자동 영역화 기법을 개발하기 위한 노력들이 계속되어 왔다. 하지만, 실제 임상적 사용에 있어 그 성능은 아직 보장되지 않는 실정이다.
이 학위 논문에서는 정지 영상 또는 동영상 등의 다양한 입력 모달리티, 단순 영역화로부터 혈관의 종류 분류에 이르기까지 다양한 문제 설정에 맞는 새로운 혈관 영역화 기법을 제안한다. 특히, 혈관의 네트워크 구조 특성 상 정형적으로 구획화된 영상 공간보다는 혈관의 구조를 반영하는 임의의 그래프 공간 상에서의 처리가 효과적일 것이라는 가설로부터 그래프 표현에 기반한 영역화 기법들을 개발하였다.
먼저, 인접 프레임 간에 혈관의 매칭점들을 찾고 이를 통해 동영상 내에서 혈관을 영역화하는 기법을 제안한다. 혈관 구조 그래프를 이용하는 이 방식은 영상 픽셀 간의 매칭점을 구하는 기존 방식과 비교하여 복잡도를 줄임과 동시에 혈관 모양의 비분별성에 따른 모호성 문제를 해결해준다. 두번째로, 그래프 신경망 (graph neural network)을 이용한 새로운 혈관 영역화 기법을 제안한다. 혈관 구조 그래프를 따라 정의된 컨볼루션 (convolution) 연산은 혈관점들 사이에 좀 더 직접적인 상호 작용을 가능하게 한다. 마지막으로, 혈관의 위상 (topology) 정보를 추정하고 이를 영역화 과정에 반영함으로써 혈관 구조를 따라 공간적으로 일관된 영역화를 수행하는 기법을 제안한다. 안저 영상과 관상 동맥 조영술 영상에서의 비교 평가를 통해 제안하는 기법들의 유용성을 확인할 수 있었다.
-
dc.description.tableofcontents1 Introduction 1
2 Vessel Segmentation via Correspondence Prediction 5
2.1 Introduction and Related Work . . . . . . . . . . . . . . . . . . . . . 5
2.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.1 Markov Random Field Representation . . . . . . . . . . . . . 8
2.2.2 Hierarchical Search of Vessel Correspondence Candidates . . 9
Global Search by Chamfer Matching . . . . . . . . . . . . . 9
Branch Search by Vessel Keypoint Correspondence . . . . . . 9
Correspondence Candidate Generation by Vessel Point Search 10
2.2.3 Post-processing for Complete Centerline Extraction . . . . . . 10
2.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.1 Evaluation Details . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.2 Quantitative Evaluation . . . . . . . . . . . . . . . . . . . . . 13
2.3.3 Qualitative Evaluation . . . . . . . . . . . . . . . . . . . . . 14
2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3 Vessel Segmentation using a Graph Neural Network 17
3.1 Introduction and Related Work . . . . . . . . . . . . . . . . . . . . . 17
3.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2.1 Overview of Network Architecture . . . . . . . . . . . . . . . 23
3.2.2 Graph Neural Network Module . . . . . . . . . . . . . . . . 24
Vertex Sampling . . . . . . . . . . . . . . . . . . . . . . . . 24
Edge Construction . . . . . . . . . . . . . . . . . . . . . . . 26
Graph Attention Network . . . . . . . . . . . . . . . . . . . 26
3.2.3 Inference Module . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2.4 Network Training . . . . . . . . . . . . . . . . . . . . . . . . 30
3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3.1 Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3.2 Evaluation Details . . . . . . . . . . . . . . . . . . . . . . . 33
3.3.3 Model Complexity . . . . . . . . . . . . . . . . . . . . . . . 35
3.3.4 Ablation Study . . . . . . . . . . . . . . . . . . . . . . . . . 36
The Effect of Vertex Sampling Density . . . . . . . . . . . . 36
The Effect of Edge Construction Methods . . . . . . . . . . . 36
The Effect of GNN Depth . . . . . . . . . . . . . . . . . . . 38
The Effect of Network Architecture . . . . . . . . . . . . . . 38
3.3.5 Comparison with Previous Methods . . . . . . . . . . . . . . 40
Quantitative Evaluation . . . . . . . . . . . . . . . . . . . . . 40
Qualitative Evaluation . . . . . . . . . . . . . . . . . . . . . 44
3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4 Vessel Segmentation with Topology Estimation 53
4.1 Introduction and Related Work . . . . . . . . . . . . . . . . . . . . . 53
4.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.2.1 Pixelwise Prediction on the Optic Disc and Vessel . . . . . . 56
Optic Disc Segmentation . . . . . . . . . . . . . . . . . . . . 56
Vessel Segmentation and AV Classification . . . . . . . . . . 56
Vascular Thickness/Orientation Classification . . . . . . . . . 58
Network Architecture . . . . . . . . . . . . . . . . . . . . . . 58
4.2.2 Graph Construction . . . . . . . . . . . . . . . . . . . . . . . 61
4.2.3 Vessel Topology Estimation . . . . . . . . . . . . . . . . . . 62
Vascular Connectivity Prediction . . . . . . . . . . . . . . . . 62
Tree Tracing via Connectivity Prediction . . . . . . . . . . . 63
4.2.4 Final AV Classification . . . . . . . . . . . . . . . . . . . . . 63
4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.3.1 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.3.2 Evaluation Details . . . . . . . . . . . . . . . . . . . . . . . 66
4.3.3 Comparison with Previous Methods . . . . . . . . . . . . . . 67
Quantitative Evaluation . . . . . . . . . . . . . . . . . . . . . 67
Qualitative Evaluation . . . . . . . . . . . . . . . . . . . . . 68
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5 Conclusion 72
Abstract (In Korean) 85
-
dc.language.isoeng-
dc.publisher서울대학교 대학원-
dc.subjectBlood vessel segmentation-
dc.subjectoptimization-
dc.subjectlearning-
dc.subjectgraph-
dc.subjecttopology-
dc.subject.ddc621.3-
dc.titleOptimization and Learning for Graph-based Blood Vessel Segmentation-
dc.title.alternative혈관 영역화를 위한 그래프 기반의 최적화 및 학습 기법 연구-
dc.typeThesis-
dc.typeDissertation-
dc.contributor.AlternativeAuthorSEUNG YEON SHIN-
dc.contributor.department공과대학 전기·컴퓨터공학부-
dc.description.degreeDoctor-
dc.date.awarded2019-08-
dc.identifier.uciI804:11032-000000157946-
dc.identifier.holdings000000000040▲000000000041▲000000157946▲-
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