Deep Learning with Class Imbalance for Detecting and Classifying Diabetic Retinopathy on Fundus Retina Images

Abstract

Abstract Diabetes mellitus is a disorder that causes diabetic retinopathy and is the primary cause of blindness worldwide. Early detection and treatment are required to reduce or avoid vision degradation and loss. For that purpose, various artificial-intelligence-powered approaches for detecting and classifying diabetic retinopathy on fundus retina images have been proposed by the scientific community. This article explores solutions to diabetic retinopathy detection by using three recently developed deep neural networks that have proven effective and efficient. Densenet201, Resnet101, and EfficientNetb0 deep neural network families have been applied to detect and classify diabetic retinopathy on fundus retina images. The dataset was notably not equilibrium; the widespread majority had been normal images, while mild Diabetic retinopathy images made up a very minor percentage of the total dataset. To treatment the skewed distribution and to keep away from biased classification results different scenarios have been used to balance the classes by utilizing (i) weight balancing with data augmentation; (ii) oversampling with data augmentation; (iii) focal loss with data augmentation, and (iv) a hybrid method of oversampling with a focal loss with data augmentation that improves the deep neural network performance of fundus retina images classification with the imbalanced dataset to build an expert system that can rapidly and adequately detect fundus images. The experimental results indicated that using Densenet201, Resnet101, and EfficientNetb0, with weight balancing on the dataset, substantially improves diabetic retinopathy prediction, by re-weighting each class in the loss function, a class that represents an under-represented class will receive a larger weight. The models yielded 94.74%, 94.74%, and 93.42%, respectively, on the test data set.