Unbiased self supervised learning of kidney histology reveals phenotypic and prognostic insights.

Deep learning methods for image segmentation and classification in histopathology generally utilize supervised learning, relying on manually created labels for model development. Here, we applied a self-supervised framework to characterize kidney histology without the use of pathologist annotations, training on whole slide images to identify histomorphological phenotype clusters (HPCs) and create slide-level vector representations. HPCs developed in the training set were visually consistent when transferred to five diverse internal and external validation sets (1,421 WSIs in total). Specific HPCs were reproducibly associated with slide-level pathologist quantifications, such as interstitial fibrosis (AUC = 0.83). Additionally, hierarchical clustering of tissue patterns revealed patient groups related to kidney function and genotype, and specific HPCs predicted longitudinal kidney function decline. Overall, we demonstrated the translational application of a self-supervised framework to summarize distinct kidney tissue patterns with phenotypic and prognostic relevance.