Multi-modal hypergraph contrastive learning for medical image segmentation

Abstract

Self-supervised learning (SSL) has become a dominant approach in multi-modal medical image segmentation. However, existing methods, such as Seq SSL and Joint SSL, suffer from catastrophic forgetting and conflicts in representation learning across different modalities. To address these challenges, we propose a two-stage SSL framework, HyCon, for multi-modal medical image segmentation. It combines the advantages of Seq and Joint SSL using knowledge distillation to align similar topological samples across modalities. In the first stage, cross-modal features are learned through adversarial learning. Inspired by the Graph Foundation Models and further adapted to our task, the Hypergraph Contrastive Learning Network (HCLN) with a teacher-student architecture is subsequently introduced to capture high-order relationships across modalities by integrating hypergraphs with contrastive learning. The Topology Hybrid Distillation (THD) module distills topological information, contextual features, and relational knowledge into the student model. We evaluated HyCon on two organs, lung and brain. Our framework outperformed state-of-the-art SSL methods, achieving significant improvements in segmentation with limited labeled data. Both quantitative and qualitative experiments validate the effectiveness of the design of our framework. Code is available at: https://github.com/reeive/HyCon.