DR-EfficientNet-L: A Distributed Deep Learning Architecture for Efficient Detection and Grading of Diabetic Retinopathy

Received: 6 July 2025 | Revised: 23 July 2025 and 3 August 2025 | Accepted: 15 August 2025 | Online: 24 September 2025

Corresponding author: Niranjan C. Kundur

Abstract

Diabetic Retinopathy (DR) is a progressive microvascular complication of diabetes and a principal cause of preventable vision loss. Existing automated DR detection systems often struggle with high computational overhead, limited generalization across datasets, and poor performance on imbalanced classes, particularly in multi-stage severity grading. To bridge these gaps, this paper proposes DR-EfficientNet-L. The framework employs a multi-branch EfficientNet backbone for parallel multi-scale feature extraction and integrates an attention-guided fusion mechanism to emphasize clinically salient lesions, such as microaneurysms and haemorrhages. A class-weighted focal loss is used to mitigate class imbalance and enhance detection sensitivity for rare severity levels. For scalability, the architecture is trained using synchronous distributed learning with gradient compression to reduce inter-node communication overhead. An extensive evaluation across EyePACS, APTOS 2019, and Messidor-2 datasets reveals superior classification performance. Under external validation, the model achieves 74.0% accuracy on Cross-Dataset (Similar) and 71.0% on Cross-Dataset (Different), reflecting robust generalization under distribution shift. The distributed setup further yields a 14.8× training speedup across 16 nodes with only 8.1% communication overhead. The comparative analysis confirms statistically significant improvements (p < 0.02) over benchmark models including DiaCNN and InceptionResNet-V2, establishing the viability of DR-EfficientNet-L for real-world, resource-aware clinical deployments.