Balancing Modality Contributions: Optimizing Fusion Mechanisms in Multi-Modal Survival Analysis
- 주제(키워드) deep learning , survival analysis , multi-modal
- 주제(DDC) 006.31
- 발행기관 아주대학교 일반대학원
- 지도교수 손경아
- 발행년도 2025
- 학위수여년월 2025. 2
- 학위명 석사
- 학과 및 전공 일반대학원 인공지능학과
- 실제URI http://www.dcollection.net/handler/ajou/000000034486
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
This thesis explores advancements in multi-modal survival analysis, addressing critical challenges in healthcare applications by integrating heterogeneous data sources such as clinical features and 3D-CT imaging. This thesis presents two innovative frameworks for multi-modal survival analysis: a modality-specific training approach for balanced and interpretable survival prediction, and a dynamic clustering framework that integrates heterogeneous data to reveal latent survival patterns. These methods address key challenges in leveraging clinical and imaging data for accurate survival analysis while enhancing interpretability in clinical contexts. The multi-modal survival prediction framework employs modality-specific encoders and a shared fusion layer to dynamically balance contributions from clinical and image data. This approach mitigates modality imbalance and ensures clinically interpretable survival predictions. Extensive evaluations on liver cancer progression and mortality prediction tasks show superior performance, achieving higher concordance indices (C-index) and lower integrated Brier scores (IBS) compared to traditional models like CoxPH and DeepHit. The multi-modal survival clustering framework extends single-modality clustering techniques, introducing modality-unified latent representations and Shapley value-based contribution analysis. This enables dynamic weighting of modalities, improving clustering and survival prediction accuracy. The approach reveals latent survival patterns, providing granular insights into diverse datasets. By addressing limitations in modality balance and interpretability, this research bridges the gap between advanced computational methods and practical clinical applications, offering actionable insights for personalized healthcare.
more목차
Chapter 1. 1
1.1 Survival Analysis 1
1.2 Multi-modal Survival Analysis 2
1.3 Modality Imbalance in Multi-modal Learning 4
1.4 Thesis Outline 6
Chapter 2. 9
2.1 Preliminaries 10
2.2 Model Architecture and Training Optimization 11
2.3 Inference and Survival Prediction 14
2.4 Data 14
2.5 Experimental Setting 16
2.6 Performance evaluation 18
Chapter 3. 25
3.1 Preliminaries 25
3.2 TCGA dataset 35
3.3 Experimental setting 36
3.4 Performance evaluation 37
Chapter 4. 44
Conclusion 44
Reference 47
