- 문제
- Oncotype DX 검사는 비싸고 접근성이 낮아, 병리 이미지(WSI)로 재발 위험을 예측하려는 시도가 많습니다. 하지만 어떤 특징 추출기와 MIL 모델 조합이 다른 기관 데이터에서도 통하는지는 체계적으로 비교된 적이 없었습니다.
- 내 역할
- 단독 연구(석사 학위논문). 데이터 정리와 WSI 전처리, 공개된 SSL·MIL 코드를 하나로 통합한 168개 조합 평가 파이프라인 구축, 실험 설계, 통계 분석, 논문 작성을 맡았습니다.
- 핵심 결과
- 외부 검증에서는 MIL 구조보다 특징 추출기 선택이 성능을 더 크게 좌우했습니다. 내부 1위였던 MIL(MeanMIL)이 외부에서는 11위로 떨어졌고, 3-class 예측에서 Low-risk 구분은 여전히 어려웠습니다.
What problem did I solve?
1. Background
Oncotype DX is expensive and not easily accessible, making treatment decisions difficult in many clinical settings.
Research Questions
- Which is more important: Feature Extractor or MIL?
- Does SSL consistently improve performance?
- Which architecture generalizes best across institutions?
Objective
Systematically evaluate SSL and MIL combinations for robust and generalizable WSI-based Oncotype DX prediction.
How did I solve it?
2. Dataset
| Cohort | Purpose |
|---|---|
| Internal | Training/Test |
| External | External Test |
| BCRNet | External Test, Cross-domain |
| TCGA | External Test, Cross-resource |
3. Method
Pipeline: Whole Slide Image → Feature Extraction → MIL → Prediction
-
Experimental Design
markdown 12 Feature Extractors × 14 MIL Models = 168 Experiments -
Evaluation strategy
markdown 3-fold Cross Validation ↓ 3 Random Seeds ↓ 9 Ensemble Models ↓ 4 Independent Test Cohorts
What were the results?
4. Results
- Finding 1
: Feature Extractor selection had a larger impact than MIL architecture under external validation. - Finding 2
: Attention-based MIL improved cross-institution generalization. - Finding 3
: SSL did not consistently outperform supervised models across all cohorts. - Finding 4
: Low-risk class remained difficult in 3-class prediction.
5. Discussion & Limitation
What was my contribution?
- Literature Review
- Dataset Curation
- WSI Preprocessing
- SSL Feature Extraction
- MIL Benchmark
- Experimental Design
- Statistical Analysis
- Visualization
- Thesis Writing
What did I learn?
Technical Insight
- Model architecture alone does not guarantee generalization.
- Representation learning plays a larger role than expected.
Research Insight
- External validation changes the ranking of models.
- Evaluation protocol is as important as model development.
Future work
- Pathology Foundation Models
- Multimodal Learning
- Better Multiclass Learning
- Clinical Integration
Appendix.
-
Attention Visualization
-
Related works
-
Feature Extractor - Supplementary
-
Multiple Instance Learning - Supplementary
-
Preprocessing: Patch Extraction
-
Metric - AUROC Range, Silhouette Coefficient
-
Information Utilization Gap
