<![CDATA[Medical image classification has advanced significantly through deep learning techniques, yet its performance remains limited by class imbalance and decentralized data silos commonly found in healthcare settings. These issues reduce model sensitivity to rare but clinically important cases, and standard Federated Learning (FL) further struggles under non-independent and identically distributed (non-IID) data. To address this, an enhanced federated model integrating unsupervised autoencoder-based reconstruction and adaptive tuning is proposed. The research contribution is an enhanced FL model that improves minority-class detection and overall classification performance under imbalanced medical image distributions, while remaining applicable across decentralized healthcare data sources. The method incorporates an autoencoder to compute reconstruction error, enabling emphasis on underrepresented samples, while adaptive tuning dynamically adjusts local hyperparameters and global aggregation weights based on sample difficulty. This integration strengthens minority-class learning without requiring additional labels or altering the decentralized structure. Experimental evaluations were conducted using two benchmark medical image datasets across three induced imbalance ratios (1:10, 1:5, 1:2) for RetinaMNIST and naturally induced imbalance for PneumoniaMNIST dataset. Results show that under severe imbalance (1:10), the enhanced model improves minority-class recall by 59.6%, F1-score by 33.9%, and AUC-ROC by 13.3% compared to standard FL. At 1:5 imbalance, recall increases by 41.3% and F1-score by 26.5%, with accuracy gains up to 6.0%. Even under mild imbalance (1:2), the model maintains consistent improvements, achieving a 26.4% recall gain and 18.9% increase in F1-score. The performance of the enhanced model was further evaluated against three baseline FL models such as standard federated learning (FedAvg), FL with GAN augmentation, and FL with standalone autoencoder-based reconstruction. The results consistently confirm that federated learning integrating adaptive tuning and autoencoder-based reconstruction outperform the three baselines FL-models for accuracy, recall, and F1-score. These findings also demonstrate that the enhanced model provides scalable, coordination-free improvement in imbalanced federated medical image classification, offering stronger performance and stability across real-world heterogeneous settings.]]>
