How doctors use similar past cases to diagnose better

Medical diagnosis relies on comparing new cases to similar historical ones, but most image classifiers ignore this. This work builds a multimodal knowledge graph from retrieved similar cases and uses graph attention to propagate diagnostic patterns into visual features. A reliability filter removes noisy case matches by jointly weighing prediction confidence and sample similarity. Tested on multiple medical imaging datasets, the method consistently beats baseline approaches and provides interpretable case-level evidence for each diagnosis decision.