Abstract: In medical image segmentation,retinal vessel segmentation is very important for the early diagnosis and treatment of ophthalmic diseases.Retinal vessel segmentation is not only helpful for the diagnosis of diseases such as diabetic retinopathy,glaucoma,and arteriosclerosis,but also has wide applications in analyzing ocular vascular morphology and hemodynamics.How-ever,existing methods cannot accurately segment small retinal blood vessels and blood vessel edges,and are still limited in terms of class imbalance,complexity of blood vessel morphology,and limited training samples.In order to improve the accuracy of blood vessel segmentation and reduce the false positive rate,this paper proposes a retinal vessel segmentation model based on multi-scale attention(MDAF-Net).The model introduces multi-scale dynamic convolution to adaptively adjust the attention to blood vessels of different scales,alleviates the problem of insufficient extraction of small blood vessels,combines channel and spatial attention mechanisms to optimize feature fusion,enhances the model’s ability to extract detailed features,and adopts a multi-scale feature fusion strategy to improve the segmentation effect under the complexity of blood vessel morphology.MDAF-Net verifies the model effect on the DRIVE and CHASE_DB1 datasets,and obtains a Dice coefficient of 0.764 and an MIoU of 78.3%(DRIVE) and a Dice coefficient of 0.820 and an MIoU of 82.5%(CHASE_DB1).The experimental results show that MDAF-Net has significant advantages in segmentation accuracy and false positive rate control,and solves the limitations of traditional methods in small blood vessel segmentation,category imbalance and false positives.

Key words: Multi-scale, Dynamic convolution, Attention fusion, Feature extraction, Retinal vessel segmentation