RetNet: Retinal Disease Detection using Convolutional Neural Network

Abstract

Retina turns light into pictures and sends messages to the brain. A retinal disease might lead to vision loss or blindness due to eye illness, ocular trauma, or other disorders. Diabetic retinopathy, AMD, and retinal detachment are some widely known retinal-based illnesses. Having eye checkup once a year can assist in preserving the health of the retina. In this matter, the utilization of machine learning and computer vision can be of significant importance. This work proposes an inexpensive, quick approach to diagnose retinal diseases correctly. In today's environment, many people utilize cellphones and high-resolution cameras and hence using computer vision to detect retinal problems will help a lot. This work proposes a lightweight custom CNN model (RetNet) to accurately diagnose and classify retinal disorders. For extensive image recognition, the convolutional neural network was fed 30904 retinal images split into 3 categories: Test, train, and validation. Four retinal conditions: CNV, DME, DRUSEN and NORMAL were deteced and classified. The CNN model trained with these datasets achieved 97.85% training accuracy and 95.41% validation accuracy. Pre-trained models such as Resnet50, InceptionV3, EfficientNetB0, Xception, and VGG16 were also used and their accuracies were 79.34%, 91.32%, 28.0%, 87.94%, and 94.01% respectively. Based on the overall research, it was clear that our lightweight custom CNN model outperformed all pretrained models and produced superior accuracy than the previous works for the used dataset.