DSC-PR: dynamic short-chain scheduling with projected relay for non-IID federated medical image classification

Federated medical image classification is impaired by label skew, missing classes, and acquisition differences across institutions, which can cause incompatible client updates and underrepresent rare categories. We propose DSC-PR, a dynamic short-chain framework that groups clients with complementary class distributions and controls conflicts during sequential model transfer. Its projected relay compares each downstream update with an exponential moving-average reference direction. When an update opposes the accumulated training direction, the conflicting component is attenuated before the model is relayed, limiting destructive cancellation while retaining locally learned information. Bidirectional traversal further reduces order bias. DSC-PR achieved a client-averaged balanced accuracy of 57.92% on Fed-ISIC2019, 4.18 percentage points above FL-DPU, and 98.21% accuracy on BloodMNIST. Ablation and diagnostic analyses showed that multi-criterion dynamic construction and projected relay jointly improved performance and optimization stability. A CPU scheduling simulation required 2.718s per round for 1,000 clients, although communication remained the main deployment cost. By reducing interference between institution-specific updates, DSC-PR may help preserve rare or center-specific imaging patterns and support more reliable multi-center medical image classification.