Fig. 2: UMAP embeddings of tibial artery tile features from three representation learning methods.

A ResNet50 with pretrained weights from ImageNet. B CTransPath (pretrained on 15 M histology images). C Our self-supervised ViT-S model trained using self-distillation with no labels (DINO). Tiles have been manually labelled with tissue substructures/pathologies to interpret clusters. DINO embeddings show both better qualitative clustering and quantitative silhouette scores, a 43% improvement over CTransPath.