Automatic classification of diabetic retinopathy through segmentation using CNN

Abstract

The process division of Diabetes Retinopathy (DR) has been considered as a significant step in diabetic retinopathy assessment and treatment. Different levels of microstructures like microaneurysm, rough exudates as well as neovascularization could take place on the retina area due to disruption to the retinal blood vessels triggered by elevated blood glucose levels. This is one of the primary causes of the prevalent visual impairment/blindness due to diabetes. Image segmentation, region merging, and Convolutional Neural Network (CNN) used in the paper for automated classification of high-resolution photographs of the retinal fundus in five stages of the DR. High heterogeneity is a significant problem for fundus image recognition for diabetic retinopathy, whereby new blood vessel proliferation including retinal detachment occurs. Therefore, careful examination of the retinal vessels is important to obtain accurate results which, through retinal segmentation could be achieved. We also highlight the difficulties in the development and learning of powerful, efficient, and reliable deep learning models for different DR diagnostic problems. The system was able to classify various DR stages with an average accuracy of around 94.2%, a sensitivity of 97%, and a specificity of 96%. There appears to be a genuine necessity for a steady interpretable classification system for DR and diabetic macular edema supported with solid confirmation. The suggested interpretable categorization systems allow diabetic retinopathy and macular edema to be properly classified. These technologies are expected to be beneficial in increasing diabetes screening and communication and discussion among those who care for these patients.