This repository contains the implementation of a Graph Laplacian Attention-Based Transformer (GLAT) integrated with an Iterative Refinement Module (IRM) for prostate cancer grading from whole-slide images (WSIs).
- IRM (Iterative Refinement Module) for adaptive patch selection, ensuring only the most informative tissue regions are selected.
- GLAT (Graph Laplacian Attention Transformer) enforces spatial coherence, preserving histological relationships.
- Convex Aggregation generates a global WSI-level representation, optimizing feature importance.
- State-of-the-art performance on five public and one private dataset.
- Computationally efficient while maintaining high accuracy.
We evaluated the model on five public and one private dataset (UConn Health Center Dataset):
| Dataset | WSIs Count | Gleason Labels | Notes |
|---|---|---|---|
| TCGA-PRAD | 895 WSIs | Gleason Grading | Public dataset from TCGA |
| SICAPv2 | 182 WSIs | Gleason Scores | High-quality annotations |
| GLEASON19 | 331 TMAs | Pixel-level Annotations | Tissue Microarrays (TMAs) |
| PANDA | 12,625 WSIs | Primary & Secondary Gleason Grades | Largest dataset used |
| DiagSet | 430 WSIs | Prostate Cancer Grading | High-quality dataset |
| Private Dataset | 79 WSIs | Clinical-grade Annotations | Internal dataset |
Patches are extracted using ([CLAM] (https://github.com/mahmoodlab/CLAM)) preprocessing pipeline:
- Stain normalization: Reduces staining variability across WSIs.
- Tissue segmentation: Removes irrelevant background regions.
- Patch extraction: Extracts 224×224 patches from WSIs.
- Filtering: Excludes patches with minimal tissue content.
To set up the environment, run:
cd ProstateCancerGrading
pip install -r requirements.txt
## 🛠 **Train the Model**
Run the following command to start training and evaluation:
```bash
python main.py