
A weakly supervised pathology pipeline that runs 368,250 tiles from 203 whole-slide images through three foundation models (UNI2-h, Virchow2, Phikon-v2) and classifies seven lung adenocarcinoma growth patterns with attention-based multiple-instance learning. Agent-selected fusion strategies reached 0.874 macro-AUROC on a one-time test set of 41 patients.
Python · PyTorch · Hugging Face · scikit-learn · W&B






























