Abstract
Pneumonia remains a serious worldwide health concern, particularly in low resource countries, where prompt diagnosis is challenging. Early detection relies on chest radiography, but data privacy rules and patient data fragmentation make AI model building difficult. Federated Learning allows collaborative model training without patient data sharing, a promising solution. Standard federated learning methods like FedAvg suffer with data heterogeneity and significant communication overhead. To overcome these constraints, this research proposes an upgraded federated framework with FedProx, which mitigates client drift in non-IID contexts by proximal optimization and Low-Rank Adaptation, a parameter-efficient fine-tuning technique that minimizes communication costs. Vision Transformers are used as the backbone architecture for chest X-ray categorization because they capture the global visual context better than convolutional models. The tiny memory footprint proposed in this research, fits resource-constrained medical infrastructure. The proposed technique was validated for a pneumonia classification job utilizing the publicly available Chest X-Ray Images dataset, which was distributed across simulated clients to replicate real-world healthcare organizations. The model's performance is measured using accuracy, precision, recall, F1-score, AUC and system-level measures including communication cost per round and convergence rate. The proposed federated model had 88.5% classification accuracy under data heterogeneity and reduced communication overhead and computation cost. Explainability research employing attention heatmaps supports the model's clinically important pulmonary areas, boosting clinical adoption, trust and transparency.