Federated Graph Learning for Pathology Images

A pathology learning pipeline that combines self-supervised image encoding, patient-level federated training, graph reasoning, and concept-bottleneck classification.

prototype Pathology Federated Learning Graph Learning Concept Bottlenecks
Source Code

project status: prototype

GLASS Path is a graph-based learning architecture for spatial structures in pathology.

The project builds a unified federated pipeline for self-supervised image encoding, local graph reasoning, and concept-bottleneck classification. It treats patients as federated clients and supports smoke-test training on constrained hardware as well as fuller training runs.

Core pieces:

  • Federated self-supervised pretraining
  • Local graph reasoning over pathology image features
  • Concept-bottleneck supervised classification
  • Patient-level dataset splitting and low-resource smoke runs