DPNet: a dual-attention patching network for breast tumor segmentation in an ultrasound image

Breast tumor segmentation plays a critical role in the early diagnosis and treatment of breast diseases. However, existing segmentation methods yet have limitations in dealing with the complex structures and low contrast of breast ultrasound images. To address this issue, we propose a novel dual-attention patching network (DPNet) for breast tumor segmentation. DPNet adopts a dual-attention structure, making efficient use of spatial information, and employs a specific fusion strategy to maximize the integration of global context and local details. To further enhance the integration of global context and local details, we introduce the spatial feature aggregation block (SFA) and the space-channel position hybrid block (SCPH). The SFA incorporates global and local attention pathways, suppressing irrelevant information and emphasizing valuable features, effectively integrating long-range spatial dependencies with fine-grained features of interest. Moreover, SCPH enhances the receptive field and improves the understanding of image structures and details by fusing spatial and channel positional information. Experimental results on the BUSI dataset demonstrate the effectiveness of our DPNet, achieving remarkable performance with Dice score of 81.10%, IoU of 73.03%, and accuracy of 96.29%. This significantly enhances the segmentation performance of breast tumor and outperforms other state-of-the-art methods in terms of segmentation accuracy and robustness.