Predicting and Classifying Diabetic Retinopathy (DR) Using 5-Class Label Based on Pre-Trained Deep Learning Models

Diabetic Retinopathy (DR) is a condition in which damage to the eyes occurs as a result of diabetes mellitus. It is the most frequent diabetes-related eye condition. It can also cause full blindness and vision loss. With effective eye treatment, the majority of new occurrences of diabetic retinopathy can be reduced. Early detection helps to avoid total vision loss. However, detecting it early can be difficult because it may not present symptoms in the early stages. The wide selection of fundus imaging makes classification challenging, mainly in Proliferative_DR, which includes the formation of new vessels in retina and bleeding. Pre-trained deep learning model is used on the publicly accessible retinal fundus image dataset on kaggle in this paper (APTOS 2019 Blindness Detection). Pre-processing and augmentation procedures are used to increase the accuracy of the models that have been pre-trained. The training accuracy of 8-Layer Convolutional Neural Network (CNN) and MobileNetV2 obtained is 83.07% and 85.21%. Testing accuracy achieved 71.93% using CNN & MobileNetV2 is 83.42%. The most often employed measures, such as the F1 Score, precision, and recall is used to ignore class level of label disagreement, which aids in diagnosing all phases of diabetic retinopathy. The results using a confusion matrix is analyzed, which is useful for categorising different stages of diabetic retinopathy according to severity. It also takes into account the degree of mismatch between the actual and anticipated labels.