DIABETIC RETINOPATHY DETECTION USING TRANSFER LEARNING BASED DEEP LEARNING MODEL
Keywords:
CNN, DL, Optimizer, Transfer learning, Augmentation, ADAM , RMSprop , SGDM, MoblieNet-V2 , GoogleNet , ResNetAbstract
The eye condition known as diabetic retinopathy (DR), which mostly affects those who have diabetes, is one of the main causes of adult blindness. Eighty percent of patients with DR have been shown to have chronic diabetes over a period of 10 to 15 years. As the disease progresses, it may result in a permanent loss of vision. Early detection allows for an early diagnosis of DR. This work focuses mostly on applying Deep Learning (DL) and transfer learning approaches to diagnose diabetic retinopathy. The Messidor 2 Dataset, which includes 1748 macula-centered eye fund images—of which 1744 are labeled and 4 are not—is taken into consideration in this study. There are two classes in the labeled data.DR0 (fundus image without diabetic retinopathy) and DR1(fundus image with diabetic retinopathy).DL models function effectively with large amounts of data. Because there isn't enough labeled data to train the Dl model, image augmentation is used. Using the principles of transfer learning, the pretrained models ResNet18, MoblieNet-V2, and GoogleNet are refined and optimized via three distinct optimizers: ADAM, RMSprop, and SGDM. These acquired deep learning models received greatest accuracy of 97.15%, 98.5%, and 98.9% for ResNet18 models, whereas models MoblieNet-V2 and GoogleNet V2 achieve lower values with less precision and performance accuracy of 74.4% and 68.1% for Adam optimizer. Our models fared better than the state-of-the-art models already in use.