Adaptive federated learning model with unsupervised autoencoder-based reconstruction for imbalanced medical image classification

Medical image classification has significantly advanced due to deep learning techniques, however, the performance remains limited by class imbalance and the nature distribution of healthcare data. Federated Learning (FL) allows collaborative model training across multiple data sources, without requiring the sharing of raw data, but its performance degrades under imbalanced and non-independent and identically distributed (non-IID) settings. This paper proposes an enhanced federated learning framework by integrating two key strategies: unsupervised autoencoder-based reconstruction, which identify and prioritises rare cases, and adaptive tuning, which dynamically adjusts aggregation behavior to balance contributions under imbalanced data conditions. These strategies improving the model’s sensitivity for underrepresented classes while maintaining stability under non-IID scenarios. The enhanced model was experimentally evaluated on two benchmark medical datasets, RetinaMNIST and PneumoniaMNIST, under three imbalance ratios (1 to 10, 1 to 5, and 1 to 2) and compared against standard FL, FL with GAN augmentation, and FL with autoencoder-based reconstruction. Results demonstrate that on RetinaMNIST, the enhanced model improved minority-class recall by 59.6 percent, F1-score by 33.9 percent over standard FL. On PneumoniaMNIST, macro recall, F1-score and AUC-ROC increased by 4.6, 4.7, and 3.3 percent respectively over the best baseline (FL with autoencoder-based reconstruction). These consistent gains indicate the framework’s scalability and robustness for federated medical image classification under real- world imbalance conditions.