MDVM-UNet:an improved multiscale-convolutional dual-pooling vision mamba model for medical ultrasound image segmentation

Abstract Accurate segmentation of plaques in carotid ultrasound images is of significant clinical value for stroke diagnosis. This paper proposes a improved UNet network based on the VMamba—MDVM-UNet. The network includes a Hybrid Convolutional 2D-selective-scan Module (HC-SS 2D) and a Hybrid Pooling 2D-selective-scan Module (HP-SS 2D) to enhance long-range modeling capabilities. HC-SS 2D incorporates multi-scale depthwise separable convolutions, improving the local details modeling. HP-SS 2D introduces a dual-pooling mechanism, reorganizing and enhancing image features. The training and testing datasets include a self-constructed carotid plaque ultrasound image dataset, as well as publicly available breast ultrasound and thyroid ultrasound image datasets. Experimental results demonstrate that MDVM-UNet achieves stronger performance across all evaluation metrics (DSC, IoU, Recall, Precision, HD95). MDVM-UNet effectively segments carotid ultrasound image plaques and improves multi-scale feature capture and long-range modeling performance.