Explainable Federated AI Framework for Privacy-Aware Medical Image Processing and Diagnostic Decision Support Systems
Abstract
Advancements in AI-driven medical imaging have improved diagnostic accuracy but often rely on centralized data collection, raising privacy, regulatory, and interpretability concerns. Existing approaches using federated learning, differential privacy, and explainable AI address parts of these issues but face challenges such as accuracy degradation, communication overhead, inconsistent explanations, and limited real-world validation. The proposed FedAvg-CNN-DPXAI framework uniquely integrates decentralized CNN training, differential privacy, and dual explainability (GradCAM and SHAP) to ensure secure, interpretable diagnostics without raw data sharing. Histopathologic Cancer Detection dataset images are preprocessed, locally trained at client sites, and aggregated via FedAvg, with model compression reducing communication costs. The framework achieves 98.2 % accuracy, 98.6 % precision, 97.5 % recall, and 97.1 % F1-score, outperforming existing models. Deployment through Docker ensures scalability and integration with hospital systems while maintaining compliance. The approach enhances diagnostic trust, preserves privacy, and supports continuous federated updates for adaptive, regulation-aligned clinical decisionmaking.