Diabetic Retinopathy Detection and Classification Using Multi-Head Self Attention Based Convolutional Neural Network

Abstract

The diabetic retinopathy is a dominant stage of diabetes mellitus which cause vision loss on the retina if it is not identified early stage. The Multi-Head Self Attention based Convolutional Neural Network (MHSA-CNN) is proposed for detect and classify the diabetic retinopathy. The Messidor and STARE dataset is used in this research and Contrast Limited Adaptive Histogram Equalization (CLAHE) and gaussian filter are used for preprocessing which enhance the contrast and removing noise. The Grey Level Cooccurrence Matrix (GLCM) is used for feature extraction and Bald Eagle Search Optimization Algorithm (BESOA) is used for feature selection. The MHSA-CNN is used for detection and classification of retinopathy which enhance the model capability at every layer without altering the parameters. The accuracy, recall, specificity, f1score and precision are used for estimating the performance. The MHSA-CNN attains accuracy 99.73% and 98.67% for Messidor and STARE dataset when compared to Modified ResNet, and Capsule Network.