From Vision to Graph Self-Supervised Learning in Digital Pathology

ICML 2025 Workshop FM4LS Submission4 Authors

Keywords: self-supervised learning, graph representation learning, computer vision, digital pathology, histopathology

Abstract: Although vision-based self-supervised learning is revolutionizing digital pathology, its domain-agnostic architectures may fail to adequately focus on the primary biological components in tissues, namely the cells and their complex interactions. We therefore propose to transform tissues into biologically informed cell graphs and investigate the effectiveness of graph SSL in encoding them. We demonstrate that pre-training on a large collection of patches using GraphMAE, with heterophilic graph neural networks, yields on par performances against popular vision-based SSL models, while using significantly fewer parameters. Finally, we show that the learned graph embeddings can effectively complement their vision-based counterparts by using a late multi-modal fusion strategy.

Submission Number: 4