A Comprehensive Deep-Learning Framework Integrating Lesion Segmentation and Stage Classification for Enhanced Diabetic Retinopathy Diagnosis

Diabetic retinopathy (DR), one of the most prevalent microvascular complications of diabetes, stands as a leading cause of vision loss globally. Due to its asymptomatic nature in early stages, delayed diagnosis and staging may result in irreversible visual impairment. Therefore, accurate and simultaneous lesion segmentation and stage classification of DR are of critical clinical importance. In this study, a two-stage, end-to-end, holistic framework is proposed for automated DR diagnosis. In the first stage, an Improved U-Net architecture enhanced with residual blocks and additional convolutional layers is employed to segment small and low-contrast lesions such as microaneurysms, hemorrhages, and hard/soft exudates with high precision. Model hyperparameters are optimized using the harmony search algorithm to enhance training efficiency. In the second stage, lesion-based weight maps obtained from the segmentation step are applied to fundus images from the APTOS dataset, generating enriched inputs for classification. A vision transformer (ViT)-based model, augmented with a Convolutional Block Attention Module (CBAM), is utilized to improve feature extraction. In addition, features derived from ViT are further refined using a graph convolutional network (GCN) and traditional machine-learning classifiers. The proposed approach achieves high performance in multi-class DR stage classification. Compared to existing studies, the framework demonstrates notable improvements in both segmentation and classification accuracy, offering a robust and generalizable solution for DR diagnosis.