Reproduce “GrapHist: Graph Self-Supervised Learning for Histopathology” on OpenResearch

About this paper

Sevda Öğüt, Cédric Vincent-Cuaz, Natalia Dubljevic, Carlos Hurtado, Vaishnavi Subramanian, Pascal Frossard, Dorina Thanou

GrapHist introduces a graph-based self-supervised learning framework that models histopathology tissues as cell graphs using masked autoencoders with heterophilic Graph Neural Networks. The framework achieves superior performance across diverse slide-, region-, and cell-level tasks, including a 5.5 percentage point improvement in tumor subtyping, while significantly reducing computational costs compared to vision transformer-based models.

Key insights

• Explicitly representing tissues as cell graphs and utilizing heterophilic GNNs provides more biologically-informed, efficient, and context-aware representations for general-purpose learning in digital pathology compared to grid-based pixel approaches.
• Large-scale self-supervised pre-training on cell graphs can generate highly transferable and robust cell-level embeddings that generalize effectively across various slide-, region-, and cell-level tasks, even in out-of-domain settings.
• Graph-based models with biological inductive biases can achieve competitive or superior performance while offering significant computational efficiencies (e.g., 2-5x fewer parameters, 3-7x faster pre-training) compared to state-of-the-art vision transformer baselines.

Summary by alphaXiv.