Accurate segmentation of vessel regions in complex arteriovenous fistula (AVF) ultrasound images, which are characterized by irregular shapes, blurred boundaries, and varied sizes, is still a significant challenge. Inspired by the remarkable performance of deep learning models in various semantic segmentation scenarios, in this paper we proposed a novel model called residual pyramidal attention UNet (RPA-UNet) for AVF ultrasound image segmentation. This model adopts several enhancements such as residual architecture network, pyramidal convolution, attention mechanism, and combined loss function, which collectively improve the model performance in terms of efficient network architecture, multi -scale feature extraction, target region feature activation, and training stability. The effectiveness of RPA-UNet has been validated through experiments on a clinical AVF ultrasound image dataset. IoU, Recall, Dice, and Precision achieved by RPA-UNet are 91.38 %, 97.21 %, 95.29 %, and 93.72 %, respectively. The results showed that the proposed model outperforms other state-of-the-art models such as Fcn32s, UNet, UNet ++, Res-UNet, and Attention-UNet. Additional experiments further prove that the enhancements of RPA-UNet contribute positively to the improvements. Thus, the proposed RPA-UNet has enormous potential for applications in complex AVF ultrasound image segmentation tasks.
