CNN-BASED DIABETIC RETINOPATHY IDENTIFICATION FROM EYE FUNDUS IMAGES

Abstract

In order to prevent patients from losing their vision and to help ophthalmologists with mass screening, automated Diabetic Retinopathy (DR) detection, screening, and diagnosis are essential. By identifying the disease before it reaches a more advanced level, DR screening seeks to treat it early. Current DR analysis techniques reduce the high cost of human computation by diagnosing using digital fundus images. In order to lessen the subjective interpretation and screening loads for ophthalmologists, researchers are persistently working toward automated screening methods. Different Convolutional Neural Network (CNN) architectures with parameter adjustment for DR classification are proposed in this research. The suggested method fine-tunes the network parameters to solve the class imbalance issue. The architecture takes into account various filter size variations, and the classification output layer analyzes how they change in response. The CNN model presented in this study has an accuracy of 87.5%, a processing time of 1 minute and 23 seconds, and a cross entropy loss of 0.6370. In fundus image categorization, the suggested strategy outperforms state-of-the-art techniques, with a maximum accuracy improvement of 13%.