Unsupervised Feature Extraction With Modern Deep Learning Approaches Determines Tissue Specific Clustering Performance

Whole-slide imaging (WSI) offers unprecedented morphological detail for digital pathology through the high-resolution digitization of histological slides. However, large-scale manual annotation is impractical due to the size, complexity, and heterogeneity of these images. Self-supervised learning (SSL) is a strategy for learning specific domain representations without labels. While this strategy is promising, comparative evaluations for unsupervised tissue clustering remain limited. We present a systematic comparison of four state-of-the-art SSL frameworks, DINO, SimCLR, DenseCL, and MoCo v2, for patch wise feature embedding in oral cavity cancer WSIs. All methods were trained under identical conditions with a ResNet-50 backbone. Features were extracted from 7,562 annotated 256x256 patches and reduced via UMAP, t-SNE, or PCA, and then clustered. Clustering performance was assessed using silhouette scores and Calinski-Harabasz indices across repeated subsampling. DINO and SimCLR consistently produced compact, well-separated clusters. These methods achieved the highest average silhouette scores (0.392 and 0.376, respectively) and Calinski-Harabasz indices (438.7 and 421.5) when paired with UMAP. DenseCL (silhouette 0.298, CH 315.5) and MoCo v2 (silhouette 0.284, CH 302.9) yielded less distinct separation and higher intra-cluster variance. These results provide practical guidance for SSL method selection in digital pathology. DINO and SimCLR showed themselves to be strong candidates for use in efficient tissue-type separation. Beyond classification, our work highlights the need to evaluate SSL representations for clustering applications and establishes a benchmark for future development of histopathology models that are optimized for clustering.