Performance Analysis of Diabetic Retinopathy Classification using CNN

Abstract

Diabetic Retinopathy (DR), a complexity induced by high blood sugar level is found to degrade the light-sensitive tissue retina by harming the blood vessels present in the region. In this work, the two models of InceptionV3 and Xception have been used as a Diabetic Retinopathy classifier to classify the given images on a ranking from 0 to 4. The APTOS 2019 dataset containing colour fundus images of various levels of severity of DR have been used to train the two models. The two models are further evaluated based on four different combinations of data pre-processing and data augmentation techniques. The Gaussian blur method was utilized for the pre-processing of the dataset. Data augmentation methods like image rotation, horizontal and vertical flips and uniform brightening were used. After comparing the performance of the two models, it was found that the Xception gave the best performance with an accuracy of 93.10% when both preprocessing and augmentation were performed on the dataset. InceptionV3 yielded an accuracy of 91.90% after employing both pre-processing and augmentation on the dataset.