Diabetic Retinopathy Automatic Detection and Classification in Fundus Images Using Modified Residual Convolutional Neural Networks (CNNs) with Improved Accuracy

The difficult task of automatically identifying and classifying diabetic retinopathy (DR) in fundus images has drawn a lot of interest from the medical imaging community. Convolutional neural networks (CNNs) have demonstrated promising performance in DR detection and other image classification tasks. The modified residual convolutional neural networks (Mod-ResNet) architecture used in this research allows the network to learn residual mappings by introducing modified residual connections. Modified ResNet has been used to identify DR and has attained state-of-the-art performance in a number of computer vision applications. In numerous DR detection and classification challenges, including the Kaggle Diabetic Retinopathy Detection Challenge, modified ResNet-based CNNs have consistently surpassed traditional machine learning methods in terms of accuracy. Because of their hierarchical and nonlinear structure as well as the automatic learning of pertinent features from raw data, CNNs exhibit higher accuracy.