Real-time Detection of Diabetic Retinopathy Using Lightweight Convolutional Neural Networks for Mobile Health Applications
Sandhya Umrao, Tripti Choudhary, Kanika Singhal, Arpita Singh, Sweety Tyagi, Arushi Goel, Vinish KumarIntroduction:
Diabetic retinopathy (DR) is a leading cause of preventable blindness worldwide. Early detection is important, though ophthalmic screening is not particularly prevalent, particularly in resource-constrained rural contexts. Deep learning and CNNs have gained popularity in automated retinal image analysis, although their high computational requirements limit their practical application in mobile health systems. This patent-oriented research aims to design an efficient, lightweight deep learning solution suitable for real-time DR detection in mobile environments.
Materials and Methods:
Retinal fundus images from Kaggle and EyePACS (~88,000 images, five DR severity classes) were preprocessed with resizing, normalization, contrast enhancement (CLAHE), sharpening, and data augmentation to balance classes. Three lightweight CNN models, MobileNetV2, ShuffleNet, and EfficientNet-B0, were initialized with ImageNet pre-trained weights and fine-tuned on these datasets using the Adam optimizer and categorical cross-entropy loss. Post-training optimizations, including pruning, quantization, and knowledge distillation, enabled deployment with TensorFlow Lite for real-time mobile inference.
Results:
EfficientNet-B0 had the best accuracy (89.6%), recall (88.9%), F1-score (89.2%), and AUC (0.957), while MobileNetV2 achieved an optimal trade-off with 87.2% accuracy, 86.4% F1, 0.934 AUC, 24 FPS, and 4.8 MB size. ShuffleNet achieved the fastest inference (28 FPS, 3.6 MB) but with lower accuracy. MobileNetV2 proved optimal for real-time mHealth deployment.
Discussion:
This study highlights the potential of lightweight CNN models in enabling real-time diabetic retinopathy detection on mobile health platforms. By employing optimized architectures, such as MobileNetV2, ShuffleNet, and EfficientNet-B0, along with TensorFlow Lite deployment, the research demonstrates cost-effective, accurate, and scalable eye-screening solutions. The findings emphasize their suitability for rural and resource-limited settings and offer significant promise for preventing blindness through early detection. The proposed patent-oriented framework underscores the novelty and applicability of this approach for future mobile health innovations and clinical deployment.
Conclusion:
This study demonstrates that lightweight CNNs can identify diabetic retinopathy in mobile hardware in real time. EfficientNet-B0 was the most precise among the models tested, while MobileNetV2 offered the best trade-off between speed, size, and accuracy. Optimised to run with TensorFlow Lite, the system offers inexpensive, ubiquitous, and on-device screening that allows early diagnosis in resource-limited health environments.