A Deep Learning Framework with Multiple Stages for Better Diabetic Retinopathy Detection

Abstract

One of the leading causes of blindness worldwide is diabetic retinopathy (DR), the most common diabetic eye disease. Retinal image classification by hand is currently a laborious procedure that calls for specific knowledge. In this paper, we looked into the possibility of automating this task using convolutional neural networks (CNNs).  A number of potent deep learning models, such as InceptionV3, ResNet50, ResNet50V2 and DenseNet201, were chosen. We used data augmentation to prevent overfitting and increase the resilience of the models. Two stages comprised our training strategy: a first round of feature extraction using a custom classifier, and a fine-tuning step where we unfroze several layers to make the models better fit our data.

We tested our approach on the PubMed, Messidor, and kaggle diabetic retinopathy dataset and discovered that it worked quite well. With a remarkable accuracy of 94.51%, the improved InceptionV3 model was the most notable.