Optical Diagnosis of Retina Healthcare with Retinia Analysis

The retinal disorders continue to be a significant factor that causes preventable vision loss, particularly when they are diagnosed at an advanced stage. Our class project created a retina analysis framework that uses deep learning to aid in screening at an early stage and in a convenient manner. In order to make the experiment fair, we used four CNN structures, ResNet50, VGG16, MobileNetV2, and EfficientNetB0, with the same experimental conditions. Our time at a glance of the performance was through accuracy trends, ROC analysis and confusion matrices. Among all the models, EfficientNetB0 was the one that provided the best balanced performance in terms of predictive power and the efficiency of the computation. As a way of rendering the black-box a bit less opaque, we used Grad-CAM visualization to show which areas of the retina affected the predictions. We packaged the model of choice into a web based application that will analyze the images in real time. The entire system is expected to be dependable, readable, and prepared to be used in practice in actual ophthalmic care institutions.