A Robust Transfer Learning Based Model To Predict Diabetic Retinopathy

Abstract

Millions of people throughout the world are affected by the disease known as diabetic retinopathy each year. As DR disorders can have a significant impact on patients' eyes and quality of life, early disease diagnosis is a crucial area of medical research. Deep Learning approaches have developed quickly for computer vision applications as a result of an abundance of data accessible for model training and improved model designs. In this paper, a strong deep-learning CNN model that performs ablation studies to classify DR illness will be described. By removing artifacts, lowering noise, and enhancing the image, the disease is colored and the overall quality is improved. The quantity of DR disease photographs has expanded because to enhancement techniques. The supplemented dataset initially proposes a base CNN model. Our suggested model, the robust CNN model, was obtained through an ablation study. A collection of freely accessible photos of DR diseases are used to train the model. With accuracy of 93.28%, the suggested robust model produced the best results. The model is solid and has excellent generalizability on new data. In cases involving the detection of binary and multi-class diabetic retinopathy diseases, the model also obtains a high level of precision and recall.