Deep-Ultranet: Diabetic Retinopathy Grading System Using Ultra-Widefield Retinal Images

Abstract

Diabetic Retinopathy (DR) is a complication due to diabetes that affects human vision. An automated and more accurate classification system is required for DR diagnosis to avoid blindness worldwide. This study presents a novel deep learning-based framework, Deep-UltraNet, designed for grading DR using Ultra-Wide Field (UWF) retinal images. The proposed system combines the strengths of dual colour space analysis (RGB and Lab) to enhance diagnostic precision. It integrates advanced preprocessing techniques, including bicubic interpolation and colour space conversion, followed by deep feature extraction through a custom Convolutional Neural Network (CNN) architecture. The custom CNN consists of four convolutional blocks using 3×3 kernels, max pooling layers, and fully connected layers for classification into four DR severity levels. The classification employs a neural network optimized with the Adam optimizer and trained via 10-fold cross-validation on the DeepDRiD dataset. The experimental results show that the proposed Deep-UltraNet provides 99.16% detection accuracy that surpasses state-of-the-art architectures such as VGG16, ResNet, and DeepUWF.