Deep learning based diabetic retinopathy screening for resource constraint applications

Diabetic Retinopathy (DR) poses a critical health concern, affecting millions of individuals globally, particularly in light of the increasing prevalence of diabetes. Early diagnosis and treatment can prevent vision loss in up to 95% of DR patients. However, the shortage of ophthalmologists, coupled with the limitations of traditional DR diagnostic methods, necessitates innovative solutions to provide timely and accessible DR screening and diagnosis, especially in underserved regions. In this study, we present a deep learning (DL) based framework for DR screening, with a specific focus on its suitability for resource-constrained edge devices. Our methodology is built upon addressing five key research questions to enhance the model’s accuracy, ensure scalability, establish robustness, optimize computational complexity, and provide a thorough performance comparison with existing models. The developed framework attains a very high accuracy of 98.4%, accompanied by a sensitivity of 95.6% and a specificity of 98.6%. It not only exhibits efficient resource utilization, with a 40-fold improvement in the accuracy-to-size ratio, but also excels with minimal CPU and GPU inference times, and low memory requirements. Our proposed framework combines enhanced accuracy and faster inference times with a lean DL model. Its adaptability facilitates inferences on edge devices in remote areas, even when cloud connectivity is intermittent. Looking forward, our focus will be on enhancing model generalization through cross-validation and hybrid databases, integrating security through federated learning, and implementing adaptive learning to ensure the model’s continued relevance in evolving healthcare scenarios. This work represents a significant contribution to the field of DR screening, providing an accessible and efficient solution for mass screening, particularly in regions with limited access to healthcare resources.