Human Knowledge Integrated Multi-modal Learning for Single Source Domain Generalization

A team of researchers has developed a breakthrough approach to one of medical AI's toughest challenges: making diagnostic models work reliably across different hospitals and imaging systems. Their system, called GenEval, combines multimodal Vision Language Models (specifically MedGemma-4B) with human medical knowledge using Low-Rank Adaptation (LoRA) fine-tuning. This addresses the critical problem of single-source domain generalization (SDG), where AI trained on data from one hospital often fails when applied to another due to subtle differences in equipment, protocols, or patient populations.

The researchers first introduced Domain Conformal Bounds (DCB), a theoretical framework to objectively measure whether domains differ in unknown causal factors. Building on this, GenEval bridges these causal gaps by integrating human expertise directly into the model architecture. In extensive testing across eight diabetic retinopathy datasets and two seizure detection datasets, GenEval achieved average accuracy improvements of 9.4% for DR grading (reaching 69.2%) and 1.8% for seizure onset zone detection (reaching 81%), significantly outperforming existing methods.

This work represents a major step toward deployable medical AI that doesn't require expensive retraining for each new hospital or imaging system. By making models more robust to domain shifts, the approach could accelerate the adoption of AI diagnostics in real-world clinical settings where data collection protocols vary widely.