CLAM

Why we included this project

CLAM is the reference implementation of a weakly supervised deep learning method for classifying whole slide images in digital pathology, published in Nature Biomedical Engineering. Teams working with digitized histopathology slides in clinical research, biobanks, or diagnostic tooling get a complete, reproducible pipeline: tissue segmentation and patching, feature extraction, attention-based multiple instance learning training, and heatmap visualization that shows which regions of a slide drove a prediction. Because it trains on slide-level labels rather than requiring patch-level annotations, it works well on real-world datasets where detailed labeling is impractical. The codebase also integrates newer pretrained histopathology encoders such as UNI and CONCH, so it stays a useful starting point even as newer tooling appears. Researchers and engineers building computational pathology workflows find a well-documented, actively maintained reference for the whole pipeline, not just a single model.