Diabetic retinal vessel segmentation algorithm based on MA-DUNet.

Background: Precise retinal vessel segmentation is crucial for the diagnosis of ocular diseases. This meticulous process can promptly capture retinal vessel abnormalities, which in turn can facilitate early diagnosis and timely treatment. However, retinal vessels exhibit complex branching patterns with different diameters and contrasts, making it difficult to segment tiny vessels and low-contrast areas, and identify subtle terminal branches. This study aimed to improve the accuracy of retinal vessel segmentation and ultimately improve the efficiency of clinical diagnosis.

Methods: In this study, we proposed the Multi-modal Attention Deformable U-shaped Network (MA-DUNet) to enhance retinal vessel segmentation performance. This model was designed to enhance segmentation accuracy through three key strategies. First, it incorporates atrous multi-scale (AMS) convolutions in the encoding stage to enhance the perception of information across different scales, thereby improving the extraction and representation of image features. Second, a gated channel transformation (GCT) attention mechanism is introduced between the encoder and decoder to improve feature transmission and extract crucial information. This mechanism dynamically adjusts the relationships among feature channels, emphasizing important task-related features while suppressing irrelevant ones. Third, a Multi-Modal Attention Fusion Block (MAFB), which combines the Multi-Modal Fusion Block (MFB) and GCT attention mechanism, is used during the decoding process to optimize information usage and enhance segmentation performance.

Results: Images from three public datasets [i.e., digital retinal images for vessel extraction (DRIVE), structured analysis of the retina (STARE), and Child Heart and Health Study in England Database 1 (CHASE-DB1)] were analyzed. In terms of blood vessel segmentation, the model had accuracy values of 95.72%, 96.68%, and 96.38%, and area under the receiver operating characteristic (ROC) curve values of 98.10%, 98.89%, and 98.35% for the three datasets, respectively. Additionally, when validated using data from hospital collaborations, our model outperformed other comparative models in terms of the segmentation results.

Conclusions: Our proposed MA-DUNet model provides more accurate segmentation of the fine terminal branches of vessels, effectively eliminating blurred boundaries and fragmentation in the segmentation outcomes, resulting in clearer vascular segmentation results.