Pathology foundation-model embeddings reveal a morphological landscape of breast lesions beyond diagnostic categories

Prov-GigaPath foundation-model embeddings map BRACS breast lesions into a train-derived morphology state space where normal-rooted kNN-Dijkstra distances recover a trajectory-like organization from normal/benign to malignant categories.

• Why it matters: Diagnostic categories overlap in morphology and reflect heterogeneous tumor pathways, making classifier-only approaches insufficient for capturing continuous morphological change across breast lesion development.
• What they did: Using the BRACS H&E dataset and a leakage-controlled, patient-aware workflow, the frozen Prov-GigaPath model generated embeddings; a train-derived PCA space supported normal-rooted kNN-Dijkstra graph-distance scoring plus order-free Energy distance and RBF-MMD2.
• The result: Validation and test projections preserved the category organization, and embeddings further dissected DCIS heterogeneity linked to concurrent invasive carcinoma without additional molecular data.

AI summary of the abstract · may contain errors · not medical advice · read the paper.