Deep CNN Architecture For Automated Identification And Severity Grading Of Diabetic Retinopathy

Authors: Vaishnavi S. Kulkarni, Suhasini A. Phatak, Sneha P. Balaki, Nikhil A. Kulkarni

Abstract: Diabetic Retinopathy (DR) is among the most prevalent microvascular diabetic complications and a major cause of avoidable blindness in the world. Early identification and correct grading of DR are critical for the initiation of prompt treatment and prevention of vision loss. Nonetheless, expert ophthalmologist-dependent retinal fundus image manual evaluation is time-consuming, prone to subjectivity, and highly reliant on skilled ophthalmologists. In order to deal with these issues, in this research, an automated, deep learning approach and Convolutional Neural Networks (CNN) is suggested for detecting the severity of Diabetic Retinopathy using retinal fundus images. The CNN architecture classifies these retinal scan images into five clinically accepted stages: NO DR, Mild Stage DR, Moderate Stage DR, Severe Stage DR, and Proliferative Stage DR. The architecture includes several convolutional layers followed by Batch Normalizationh,aRrdeLU Aectxiuvadtaiotens,, and Msoaftx Pooleinxugdoapteesr,ationsantod extract hierarchical features of retinal abnormalities. nAeFouvlalysccuolanrnizeacttieodn.layer approach is used to avoid overfitting and improve generalization. The last softmax layer gives probabilistic classification output. The model is learned and being tested on a vast annotated dataset of retinal images with high classification accuracy and sensitivity for all DR grades. Experimental results have proven that the suggested CNN model efficiently detects the DR stages with great reliability. The system provides a scalable and non-invasive method to aid ophthalmologists in early screening and diagnosis, thus facilitating better treatment, particularly for resource-poor areas.

DOI: 10.61463/ijset.vol.13.issue3.184