728 regions of interest extracted from pathological images using QuPath. Three classes: Tubular Adenoma, GVHD (graft-versus-host disease), and None. Feature extraction via Prov-Gigapath foundational model (1x1536 feature vectors).
ROIs extracted at 128x128 and 256x256 pixels, approximately 40-50 per image.
| Metric | Test | Cross-val |
|---|---|---|
| AUC | 0.982 | 0.944 |
| Accuracy | 0.877 | 0.86 |
| Precision | 0.9 | 0.874 |
| Recall | 0.854 | 0.849 |