Highly Efficient On-Device Diabetic Retinopathy Screening Using Quantized EfficientNet-Lite Models on Retinal Fundus Images

This research proposes a highly efficient, quantized convolutional neural network framework for on-device diabetic retinopathy (DR) screening, integrating lightweight EfficientNet-Lite models, lesion-preserving image preprocessing, and INT8 quantization for real-time edge deployment. Using EfficientNet-Lite0 as the architectural backbone and applying adaptive histogram equalization and vessel segmentation, the system was benchmarked across clinical-grade fundus image datasets including EyePACS, APTOS 2019, and Messidor-2. The model consistently achieved DR detection accuracies up to 91.3%, with F1-scores reaching 0.859 and inference times below 130ms on ARM-based mobile processors, validating its diagnostic precision under resource constraints. Compared to unquantized baselines and mobile CNNs, the proposed framework demonstrated a 4.7–5.2% increase in accuracy and a 45% reduction in memory footprint, with no significant loss in classification performance due to quantization. TensorFlow Lite optimization and lesion-centric preprocessing enabled consistent screening performance under low-light and noisy imaging environments. With the compressed model size of ~4.2MB and readiness for deployment onto real-world smartphone and embedded systems, the proposed pipeline facilitates scalable, privacy-friendly, and accurate DR screening in remote or underserved locations.