SGF-SCA: A Spatial Gated Framework with Shared Channel Attention for Breast Ultrasound Image Segmentation

Tianyu Xu, Yue Li, Zixiang Song, Yipeng Xu, Yuzhe Wu, Xiangjian He

Abstract: Breast ultrasound (US) image Segmentation is a significant challenge due to high noise levels, blurred boundaries, and low signal-to-noise ratios. To address these issues, we propose SGF-SCA, a Swin-Unet-based architecture enhanced with Spatial Gated Unit (SGU) and Shared Channel Attention Fusion Unit (SCAFU). The SGU in the encoder focuses on suppressing irrelevant features and enhancing relevant tokens in US images. In the decoder, SCAFU efficiently fuses skip connections with up-sampled features using shared channel attention mechanism. The enhancements improve feature extraction and refinement and facilitate solving the above issues of US image segmentation. Our method demonstrates improved performance on the BUSI dataset, with a Dice Coefficient of 79.9%, showing robustness in handling the complexities of ultrasound images.