DINOv2-KMeans: Self-Supervised Tissue Subtype Discovery from Histopathology Patches Using Vision Transformer Embeddings and Unsupervised Cluster Optimisation

Abstract

The automated discovery of clinically meaningful tissue subtypes from histopathological images remains a fundamental challenge in computational pathology, hampered by the prohibitive cost and scarcity of expert annotations at scale. This paper proposes DINOv2-KMeans  a self-supervised tissue subtype discovery framework that extracts rich 768-dimensional patch-level embeddings from the DINOv2 Vision Transformer (ViT-Base/14, pretrained via self-distillation with no labels on 142M images) and applies PCA-compressed K-Means clustering with principled cluster count optimisation to discover biologically coherent tissue subtypes without any label supervision. Evaluated on PathMNIST  a standardised public proxy for multi-tissue histopathological classification comprising 107,180 colon pathology patches from 9 tissue classes — with N=10,000 sampled patches, DINOv2-KMeans discovers K=5 optimal subtypes achieving Normalised Mutual Information (NMI) of 0.9312, Adjusted Rand Index (ARI) of 0.9189, Cluster Purity of 0.9923, and Silhouette Score of 0.5480. These results substantially exceed all eight compared state-of-the-art self-supervised tissue discovery methods by a minimum of 3.0 NMI percentage points. PCA dimensionality reduction (768-D → 128-D, retaining 99.23% of variance) precedes clustering, with optimal K selected by multi-criterion ranking of Silhouette, Calinski-Harabasz, and Davies-Bouldin indices across K=2 to 10. Post-hoc UMAP and PCA-2D visualisations confirm compact, well-separated embedding structures aligned with ground truth tissue categories. DINOv2-KMeans establishes a new state-of-the-art for annotation-free tissue subtype discovery and provides a reproducible, scalable foundation for AI-assisted pathological tissue atlasing.