Skip to content

IMPROVED_PLAN.md

Latest commit

 

History

History
61 lines (51 loc) · 2.92 KB

IMPROVED_PLAN.md

File metadata and controls

61 lines (51 loc) · 2.92 KB
Raw

Improved Execution Plan: CLAM vs. TransMIL Benchmarking

This plan outlines the updated protocol for a comparative study between CLAM and TransMIL architectures, optimized for the current project state and hardware constraints (NVIDIA 1650, 4GB VRAM).

1. Research Objective

Benchmark the performance of CLAM (Single-Branch) and TransMIL on the task of 5-year survival classification using Whole Slide Image (WSI) features from the TCGA-BRCA cohort.

2. Experimental Setup

  • Feature Extractor: Fixed (UNI2-h embeddings).
  • Dataset: TCGA-BRCA (Breast Invasive Carcinoma).
  • Target Variable: Binary 5-year survival (label: 1 if deceased < 1825 days, 0 otherwise).
  • Validation Strategy: 5-fold cross-validation (Stratified by Patient/Case ID).
  • Validation Split: 20% of training data in each fold (increased from 10% for evaluation stability).

3. Data Infrastructure

  • Feature Directory: TCGA-BRCA_IDC/ (H5 format).
  • Master Manifest: master.csv (contains slide_id, case_id, and label).
  • Splits: splits/ (Generated using generate_splits.py). Each fold directory (0-4) contains train.csv, val.csv, and test.csv.

4. Phase I: CLAM Implementation

CLAM serves as the attention-based MIL baseline.

  • Script: CLAM/main.py
  • Configuration (Optimized for Overfitting):
    • model_type: clam_sb
    • bag_loss: focal
    • bag_weight: 0.8 (Balanced clustering emphasis)
    • lr: 3e-5
    • reg: 1e-3 (Aggressive weight decay)
    • drop_out: 0.5
    • early_stopping: Enabled
  • Execution Command (Example Fold 0): 
    

    • 



python main.py --data_root_dir "D:/thesis project/TCGA-BRCA_IDC" --split_dir "D:/thesis project/splits" --task task_2_tumor_subtyping --exp_code brca_5yr_classification_clam_final --model_type clam_sb --bag_loss focal --lr 3e-5 --k 5 --k_start 3 --k_end 5 --subtyping --B 8 --drop_out 0.5 --reg 1e-3 --bag_weight 0.8 --early_stopping



## 5. Phase II: TransMIL Implementation
TransMIL serves as the transformer-based challenger.
- **Script:** `TransMIL/train.py`
- **Configuration:** `TransMIL/TCGA_BRCA.yaml`
- `optimizer.lr`: `2e-5`
- `loss.base_loss`: `focal`
- `epochs`: `200`
- `weight_decay`: `0.01`
- **Execution Command (Example Fold 0):**
```bash
python TransMIL/train.py --stage 'train' --config 'TransMIL/TCGA_BRCA.yaml' --gpus 0 --fold 0



6. Phase III: Evaluation & Statistical Analysis


Compare models based on:


    

  • Area Under the ROC Curve (AUC): Primary metric for classification.
    • 

  • Concordance Index (C-Index): Assessing risk ranking.
    • 

  • DeLong's Test: Statistical significance of AUC differences.
    • 

  • Log-rank Test: Comparing survival curves of high-risk vs. low-risk groups.
    • 



7. Phase IV: Interpretability


Generate attention heatmaps to visualize the regions of interest for both models using the scripts in CLAM/create_heatmaps.py (adapted for TransMIL attention weights).




Last Updated: March 2026