DMANet: A Medical Ultrasound Image Segmentation Network Based on Dual-Stream Multidimensional Attention

Due to the diversity of object types and scales, as well as the similarity in appearance between background tissues, extracting valuable information from different medical images is challenging. In this paper, we introduce a dual-stream multidimensional attention network (DMANet) for medical image segmentation. DMANet is composed of dual-stream pyramid modules, an encoder-decoder module with hybrid multidimensional attention, and a multi-scale feature fusion module for the final output. Specifically, the dual-stream pyramid captures features at different scales and retains many useful original details by processing inputs of different resolutions from the original images, facilitating the learning of local detailed features at various scales. The skip connection module with Multi-dimensional Attention Gate (MDAG) suppresses complex and irrelevant information in both channel and spatial dimensions, while learning richer contextual information. Additionally, to address the issue of limited receptive fields, we propose a multi-scale feature fusion block (MFF), achieving higher segmentation accuracy. We validated the superiority of our network on two public datasets by comparing it against eight benchmark models.