A Mobile Feature-Based Cascaded Deep Learning Approach for Diabetic Retinopathy Classification

Diabetic Retinopathy is a serious eye condition primarily caused by impaired blood vessels at the posterior part of the eye. This condition can vary in severity, ranging from mild to proliferate, and its progression can cause impaired vision, leading to complete blindness. This paper proposes a cascaded deep-learning framework with a mobile fundus scanner to classify diabetic retinopathy through feature space enhancement and a cascade of three convolutional neural network models. Each model focuses on a specific classification task, activating the subsequent model if further classification is required. Our methodology initially involved establishing three baseline models for binary classification, multi-class classification, and classification of the mild and moderate classes through transfer learning of five state-of-the-art models, namely Xception, ResNet50V2, InceptionV3, DenseNet201, and EfficientNetV2S. We then explore two feature space enhancement techniques, namely unsharp masking and fuzzy contrast enhancement, the latter of which addresses the limitations of conventional histogram equalization techniques. Following model evaluation, EfficientNetV2S was selected as the base model for all three classifiers in the cascaded framework, resulting in accuracies of 87%, 91%, and 83% for the binary, multi-class, and mild/moderate classifiers, respectively. The results show that our proposed framework exhibits an enhanced performance in classifying stages that share similar features, such as the mild and moderate stages, as well as the severe and proliferate stages, outperforming the performance of published related work and highlighting areas for future research that can open the door for the early detection of diabetic retinopathy.