IoT-Enabled Sustainable Deep Learning Framework for Multi-Class Ophthalmic Disease Detection using Fundus Images

This paper presents a smart IoT-embedded ophthalmic screening platform using novel deep learning techniques to automate and to enhance the resource management for eye care. Connected imaging devices capture fundus images and the images are pre-processed (normalized and augmented) at the network edge. The feature extraction is implemented through a pre-trained CNN-ResNet-50 and InceptionV3 using hyper parameters fine-tuned to classify multiple retinal diseases (cataract, glaucoma, retinal tear, and normal) in one pipeline. Screening results are sent wireless to a central cloud-based dashboard, which allows for remote monitoring and real-time triage of diagnostic resources (such as specialist-based referral, and load balancing across clinics). Experimentation based performance on accuracy, precision, recall and F1score asserts the credibility of the proposed system. Because of the all-in-one multi-class detection in an IoT native infrastructure, the system supports scalable teleophthalmology, reduces specialist burden, and improves resource utilization all of which are essentials towards sustainable health delivery in the under-served areas.