Applying Self-Supervised Tissue Patch Encoders to Enable Cellular Graph Encoding

Recent advances in computation pathology have seen the development of various forms of foundational models that have enabled high-quality, generalpurpose feature extraction from tissue patches. However, most of these models are somewhat limited in their ability to capture cell-to-cell spatial relationships essential for understanding tissue microenvironments. To address this issue, self-supervised learning (SSL) applied to cellular graph models is a promising avenue of investigation as this would afford one the ability to capture cell-level interpretability of the tissue microenvironment and learn relationships from the cellular topology. Compared to SSL on tissue patches, developing cellular graph models requires considerable computation and algorithmic effort, starting with accurate cell detection/segmentation from tissue and then subsequently requiring adequate cell-level feature extraction. In this study, we demonstrate that a pretrained tissue patch encoder, originally trained in the context of the token used in conventional vision transformer frameworks can also serve effectively as a cellular graph encoder for cellbased patches . We report three key findings: 1) the pretrained tissue patch encoder acts as a powerful graph-level feature extractor for cellular graphs in tissue patches; 2) it also can function as a celllevel feature extractor, allowing each cell to be contextualized within its local tissue microenvironment environment; 3) the cell patch size and positional embeddings are critical factors for high-quality cellular graph feature extraction. To validate and support these insights, we evaluate our approach on various tissue region classification tasks from three different types of cancers. Experimental results demonstrate that our method achieves performance comparable to the current state-of-the-art foundation models in computational pathology. These results highlight the potential of pretrained tissue patch encoders in cellular graph modeling, which can enable insights into how the cellular features and topology are related to various biological processes and phenomena.