One-Round Federated Learning for Medical Imaging

Federated learning (FL) offers a privacy-preserving way for multiple medical institutions to collaborate on training AI models without sharing sensitive patient data. However, a major challenge arises from "task heterogeneity," where each institution only labels a subset of diseases, leaving other categories unobserved. Traditional gradient-based FL methods struggle here due to slow convergence and systematic false-negative biases from missing labels. This research proposes a novel analytic federated learning framework specifically designed for multi-label medical image classification under such task heterogeneity. The method replaces iterative gradient optimization with three closed-form operations: a balanced label projection to correct class-imbalance bias, a per-class absolute aggregation that optimally combines ridge-regression classifiers from contributing clients, and an optional analytic pseudo-label refinement. This entire process remarkably requires only one or two communication rounds, regardless of the number of clients or the degree of task heterogeneity. Experiments on the ChestXray14 dataset demonstrate substantial performance improvements over state-of-the-art federated multi-label methods, significantly boosting accuracy and AUC while drastically reducing communication overhead.