DPSMUNet: a new network based on a dual-pooling self-attention module for carotid artery plaque segmentation in ultrasound images

Plaque segmentation in ultrasound images is essential for diagnosing carotid artery diseases but is hindered by challenges like noise, low contrast, and irregular plaque shapes. This paper proposes DPSMUNet, an enhanced UNet-based deep neural network incorporating a Dual Pooling Self-Attention Module (DPSM) and a Depthwise Separable Feedforward Network (DSFN). DPSM enables multi-scale global self-attention, while DSFN fuses semantic information across layers. Our experimental results, conducted on our carotid ultrasound image dataset as well as publicly available datasets (breast ultrasound image dataset and nerve ultrasound image dataset), demonstrate that DPSMUNet achieves strong performance across all metrics (DSC, IoU, recall, precision, parameter count, and inference time), with DSC, IoU, and recall being the most optimized. These findings suggest that DPSMUNet effectively segments plaque boundaries in carotid ultrasound images with relatively low computational complexity.