Diabetic Retinopathy (DR) is a prominent microvas-cular complication of diabetes and a leading cause of preventable blindness worldwide. While Deep Learning, specifically Convolu-tional Neural Networks (CNNs), has established itself as the gold standard for automated retinal screening, the centralization of sensitive medical imagery for model training violates stringent data privacy regulations (e.g., HIPAA, GDPR, DPDP). Further-more, practical clinical deployment requires grading severity rather than simple binary detection. This paper presents a highly comprehensive, privacy-preserving distributed computing framework and an integrated Clinical Decision Support System (CDSS) designed for 5-class multi-label DR diagnosis. Leveraging Federated Learning (FL), the architecture empowers heteroge-neous healthcare institutions to collaboratively train a modified ResNet-50 architecture without transferring raw patient data. We address the critical bottleneck of communication overhead and non-IID (Independent and Identically Distributed) clinical data distributions by introducing a partial fine-tuning strat-egy—freezing all but the final 15 layers of the ResNet-50 back-bone. Evaluated on the APTOS 2019 Blindness Detection Dataset (3,662 retinal fundus images) distributed across 10 simulated clinical clients, the federated model’s efficacy is tested under both IID and Non-IID (Quantity and Label Skew) conditions. The

proposed model achieved an exceptional accuracy of ∼99.00% in IID settings and demonstrated high resilience by maintaining

∼96.00% accuracy under severe Non-IID quantity and label skew constraints. To bridge the gap between theoretical modeling

and clinical application, the system incorporates a full-stack Flask web application featuring robust user authentication and real-time diagnostic analytics. This end-to-end system provides a scalable, privacy-compliant solution for comprehensive DR grading, effectively eliminating raw data leakage risks.

Index Terms—Diabetic Retinopathy, Federated Learning, Dis-tributed Computing, ResNet-50, Non-IID Data, Partial Fine-Tuning, Privacy-Preserving AI.