KatherLab/STAMP

Implements weakly-supervised multiple instance learning with Transformer aggregation to train biomarker models directly from slide-level labels, eliminating the need for pixel-level annotations. Supports 20+ foundation models (Virchow-v2, UNI-v2, TITAN, COBRA) for feature extraction and enables classification, multi-target classification, regression, and Cox survival analysis through a unified YAML-configured CLI. Scales from local execution to HPC clusters (SLURM) with built-in explainability via attention heatmaps and tile exports, validated across multi-center tumor cohorts.

115 stars.