Abstract
Computer-aided diagnostic (CAD) systems for color fundus images play a critical role in the early detection of fundus diseases, including diabetes, hypertension, and cerebrovascular disorders. Although deep learning has substantially advanced automatic segmentation techniques in this field, several challenges persist, such as limited labeled datasets, significant structural variations in blood vessels, and persistent dataset discrepancies, which continue to hinder progress. These challenges lead to inconsistent segmentation performance, particularly for small vessels and branch regions. To address these limitations, we propose an enhanced HarDNet-based model that integrates HarDNet modules, Receptive Field Block (RFB) modules (designed to capture multi-scale contextual information), and Dense Aggregation modules. This innovative architecture enables the network to effectively extract multi-scale features and improve segmentation accuracy, especially for small and complex structures. The proposed model achieves superior performance in retinal vessel segmentation tasks, with accuracies of 0.9685 (±0.0035) on the DRIVE dataset and 0.9744 (±0.0029) on the CHASE_DB1 dataset, surpassing state-of-the-art models such as U-Net, ResU-Net, and R2U-Net. Notably, the model demonstrates exceptional capability in segmenting tiny vessels and branch regions, producing results that closely align with the gold standard. This highlights its significant advantage in handling intricate vascular structures. The robust and accurate performance of the proposed model underscores its effectiveness and reliability in medical image analysis, providing valuable technical support for related research and applications.