RESOURCE-EFFICIENT DEEP LEARNING FOR FIVE-STAGE DIABETIC RETINOPATHY CLASSIFICATION: COMPARATIVE EVALUATION, MODEL OPTIMISATION, AND EDGE DEPLOYMENT

Abstract

Diabetic retinopathy (DR) is a major cause of preventable vision impairment, particularly in regions where access to ophthalmologists and computational infrastructure is limited. This study investigates resource-efficient deep learning for automated five-stage DR classification from retinal fundus images, with emphasis on balancing diagnostic performance and deployment feasibility. Five convolutional neural network architectures—EfficientNet-B0, MobileNetV2, ResNet50, DenseNet121, and a custom lightweight CNN—were evaluated using 3,662 fundus images categorized as no DR, mild non-proliferative DR (NPDR), moderate NPDR, severe NPDR, and proliferative DR. Images were preprocessed using retinal-region masking, contrast enhancement, normalization, and data augmentation, while weighted loss was employed to mitigate class imbalance. Model performance was assessed using balanced accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve, together with model size, parameter count, computational complexity, memory consumption, and inference latency. EfficientNet-B0 achieved the highest balanced accuracy of 86.4% and an AUC-ROC of 0.957, whereas MobileNetV2 achieved 82.7% balanced accuracy with substantially lower computational requirements, requiring only 9 MB of storage and 21 ms inference time. For early-stage DR analysis, EfficientNet-B0 obtained 85.2% sensitivity and 92.1% specificity. Model pruning and quantization further improved deployment efficiency, with combined optimization reducing model size by up to 85%, although greater compression resulted in moderate performance degradation. Hardware-based evaluation on edge, mobile, and CPU-based platforms demonstrated that optimized MobileNetV2 provided the most favorable balance between predictive performance, memory requirements, latency, and energy consumption. These findings demonstrate that appropriately optimized lightweight deep learning architectures can support computationally efficient DR grading and provide a practical foundation for AI-assisted retinal screening in resource-constrained healthcare environments.