mahmoodlab/CLAM

Implements attention-based multiple instance learning (CLAM) to classify whole slide images using only slide-level labels, automatically identifying diagnostically valuable sub-regions without patch-level annotations. The pipeline integrates pretrained encoders like UNI and CONCH for feature extraction, then applies clustering-constrained attention to refine the feature space and enable interpretable heatmap visualizations of diagnostic regions. Handles multi-class histopathology tasks across diverse WSI formats (.svs, .ndpi, .tiff) with cross-domain generalization to different tissue types and acquisition modalities.

1,625 stars. No commits in the last 6 months.