Contextual unsupervised deep clustering in digital pathology

Clustering can be used in medical imaging research to identify different domains within a specific dataset, aiding in a better understanding of subgroups or strata that may not have been annotated. Moreover, in digital pathology, clustering can be used to effectively sample image patches from whole slide images (WSI). In this work, we conduct a comparative analysis of three deep clustering algorithms - a simple two-step approach applying K-means onto a learned feature space, an end-to-end deep clustering method (DEC), and a Graph Convolutional Network (GCN) based method - in application to a digital pathology dataset of endometrial biopsy WSIs. For consistency, all methods use the same Autoencoder (AE) architecture backbone that extracts features from image patches. The GCN-based model, specifically, stands out as a deep clustering algorithm that considers spatial contextual information in predicting clusters. Our study highlights the computation of graphs for WSIs and emphasizes the impact of these graphs on the formation of clusters. The main finding of our research indicates that GCN-based deep clustering demonstrates heightened spatial awareness compared to the other methods, resulting in higher cluster agreement with previous clinical annotations of WSIs.