Improving Diagnosis and Classification of Severity in Diabetic Retinopathy Using Transfer Learning

One of the most severe consequences of diabetes mellitus is diabetic retinopathy, which causes degeneration of blood vessels in the retina and blindness. The deadly effect of the described disease can be potentially prevented or addressed if detection and treatment measures are undertaken timely. The recent implementation of deep learning (DL) approaches in analyzing massive and complicated data has had a transformative effect of the sphere of medical image analysis. Being capable of recognizing the tiniest features and specifics of the retina, these methods enable highly complicated, efficient detection and classification algorithms. This paper provides a detailed description of the methodology for classifying the level of DR severity using retinal scans from Kaggle. The timeliness of the above response is explained by the urgency of DR detection and treatment measures. It should be noted that a specifically arranged data set was created in the described work, including 2750 pictures in with 5 groups of severity. The maximum training accuracy of 98.55% was reached by using EfficientNetB3 with transfer learning, and the validation and testing accuracies constituted 71.27% and 76.36%, respectively.