Diabetic Retinopathy Classification Based on Segmented Retinal Vasculature of Fundus Images Using Attention U-NET

Abstract

Retinal fundus image analysis (RFIA) can identify diabetic retinopathy (DR) by which diabetic patients can lower the chance of losing vision. Applying profound learning on clinical information is an exceptionally tedious task. In this research paper, automatic detection and classification of DR using various deep learning networks are presented. Here, special emphasis is given for detection of diabetic retinopathy based on segmented retinal vasculature of fundus images using Attention U-NET. The experimentation is done with the use of three different dataset-Indian Diabetic Retinopathy Dataset (IDRD), the Handy Aptos 2019 Blindness Detection Database, and a dataset from Kaggle. Experimental analysis was also done with both imbalanced and balanced dataset. We utilized several deep learning models including VGG19, ResNet50 and DenseNet201 to find the best arrangement model for the DR location. In order to address the inadequacies of these models, we used three more U-Net models: Simple U-NET, Res-UNET, and Attention U-NET. The model results have been examined using a number of measures including precision, recall, F1-score, and accuracy. The proposed model’s highest level of accuracy achieved is 95.6%. Performance of our proposed methodology is compared with that of recent similar works and found to be superior. The extensive simulation studies reveal the suitability of the suggested strategy for clinical application.