EdgeRetina: Hybrid multimedia architecture for diabetic retinopathy screening on low-cost mobiles

Abstract

Diabetic retinopathy (DR) is a major cause of preventable blindness, particularly in areas with limited medical resources where access to ophthalmologists is critical. Existing automated solutions struggle to balance clinical performance, cost-effectiveness, and robustness in the face of fundus image variability—including lighting differences, artifacts, and uneven capture quality. To address this challenge, we propose EdgeRetina, an integrated solution for diabetic retinopathy screening on low-cost mobiles. Our approach combines lightweight preprocessing (128×128 resizing, intensity normalization, and targeted augmentations simulating real-world conditions) with a hybrid SqueezeNet-MobileViT architecture (1.4 million parameters), optimized by dynamic threshold calibration (median: 0.3), maximizing clinical utility. Clinically calibrated INT8 quantization reduces the model to 8.27 MB (-92%) without altering diagnostic performance (sensitivity of 90.7% for referable diabetic retinopathies), while preserving compatibility with floating point 32 (FP32)-based gradient-weighted class activation mapping (Grad-CAM) visualizations. Evaluated on the APTOS 2019 dataset, this solution achieves an AUC of 0.96 with a latency (inference time) of 15.43 ms, reducing CPU consumption by 43% compared to FP32. The dynamic threshold/INT8 coupling decreases false positives by 71.4%. This pipeline thus enables accurate, accessible, and early screening of diabetic retinopathy on low-cost mobile devices, combining operational efficiency and diagnostic reliability in constrained environments, which is crucial to prevent avoidable blindness.