Deep Learning Approach for Retinopathy Identification on Combined Clinical Datasets

Abstract—This work presents a system for automatic detectionof various stages of diabetic retinopathy (DR) based on fundusimages of patients. The system was built based on a relatively newand little-used image database: ”Dataset of fundus images for thestudy of diabetic retinopathy” version v3 CastilloBenitez21. Theprimary dataset was expanded using clinical fundus photographsacquired from the Department of Nephrology at Wroclaw MedicalUniversity. The diagnostic system was developed based onvarious variants of convolutional neural networks (CNNs) thatwere pre-trained on ImageNet data. The CNN classifier, based onVGG16 with transfer learning, proved to be effective and gavea global accuracy of 83.15%. The evaluation of discriminationbetween the non-DR and the DR state resulted in an accuracyof 89.7%, with a sensitivity of 94.9%, a specificity of 88.3%, anda Matthews Correlation Coefficient of 0.7665.