DGIST Scholar: Weakly Supervised Representation Learning for Histopathology Image Analysis

Department of Robotics and Mechatronics Engineering Theses Ph.D.

Cited time in webofscience

Cited time in scopus

Weakly Supervised Representation Learning for Histopathology Image Analysis

Title: Weakly Supervised Representation Learning for Histopathology Image Analysis

Alternative Title: 조직병리학 이미지 분석을 위한 약지도표현학습

Author(s): Philip Chikontwe

DGIST Authors: Philip Chikontwe ; Sang Hyun Park ; Sunghoon Im

Advisor: 박상현

Co-Advisor(s): Sunghoon Im

Issued Date: 2023

Awarded Date: 2023-02-01

Type: Thesis

Description: Computer aided diagnosis, Deep learning, Histopathology, Multiple instance learning, Weak supervision

Table Of Contents: I. Introduction 1
1 Background and Motivations 1
1.1 Weakly Supervised Learning 3
1.2 Multiple Instance Learning 4
1.3 Histopathology Image Analysis 6
2 Contributions and Outline 7
3 Publications 8
3.1 Excluded Research 9

II. End-to-End Multiple Instance Learning with Center Embeddings 11
1 Introduction 11
2 Related Works 12
2.1 Multiple Instance Learning for Histopathology 13
2.2 Unsupervised Methods with Multiple Instance Learning 13
3 Methodology 13
3.1 Top-k Instance Selection 14
3.2 Instance Level Learning 14
3.3 Pyramidal Bag-level Learning 15
3.4 Soft-Assignment based Inference 15
4 Evaluation 16
4.1 Dataset and Settings 16
4.2 Quantitative Results 17
4.3 Qualitative Results 18
5 Conclusion 19

III. Dual Multiple Instance Learning with Self-Supervision 20
1 Introduction 20
2 Related Works 22
2.1 Deep Learning for COVID-19 Diagnosis 22
2.2 COVID-19 Diagnosis under Weak Supervision 22
2.3 Self-Supervised Learning 23
3 Methodology 23
3.1 Dual Attention based Learning 25
3.2 Contrastive Multiple Instance Learning 26
4 Evaluation 27
4.1 Datasets and Settings 27
4.2 Quantitative Results 29
4.3 Qualitative Results 30
5 Discussion 31
6 Conclusion 32

IV. Weakly Supervised Segmentation in Gigapixel Images 33
1 Introduction 33
2 Related Works 35
2.1 Weakly Supervised Segmentation in Computer Vision 35
2.2 Patch-based Weakly Supervised Segmentation for Histopathology 36
2.3 Weakly Supervised Learning with Self-Training 36
3 Methodology 37
3.1 Contrastive Neural Compression 38
3.2 Self-Supervised Neural CAM Refinement 39
3.3 τ -MaskOut: Spatial Feature Masking as Regularization 40
3.4 Pixel Correlation Module (PCM) 41
3.5 Conditional Entropy Minimization between CAMs 42
4 Evaluation 43
4.1 Datasets 43
4.2 Implementation Settings 44
4.3 Quantitative Results 46
4.4 Ablations on Learning Objectives 49
4.5 Qualitative Results 50
4.6 Ablations on Relative Tumor Sizes 51
5 Conclusion 52

V. Re-calibrating Feature Distributions in Histopathology 53
1 Introduction 53
2 Methodology 55
2.1 Preliminaries: A Simple Baseline 56
2.2 Feature Re-calibration & Max-Instance Selection 57
2.3 Positional Encoding Module (PEM) 57
2.4 MIL Pooling by Multi-head Self-Attention (PMSA) 58
3 Evaluation 59
3.1 Datasets and Settings 59
3.2 Main Results 59
3.3 Impact of Learning Objectives 61
4 Conclusion 61

VI. Concluding Remarks 62
1 Conclusion 62
2 Future Work 63

VII. Acknowledgement 64
References 65