Abstract:Automatic segmentation of breast ultrasound images is crucial for improving the diagnostic accuracy of breast cancer. Although Convolutional Neural Networks (CNNs) and Transformers have achieved remarkable advancements in medical image analysis, they face specific challenges when applied to lesion segmentation. Notably, the iterative processes of downsampling and self-attention computation excel at extracting significant features from both local and global perspectives. However, these processes inadvertently compromise the spatial properties of pathological regions and disrupt the reconstruction of lesion boundaries. In response to this challenge, this study proposes a novel Global Boundary Enhancement (GBE) mechanism for the segmentation of breast ultrasound images, named GBE-Net. Specifically, a component named GBE-Block is initially designed to learn the spatial and boundary features of ultrasound images. Then, based on the GBE-Block and an edge detection operator, a double-branch encoder is constructed, which incorporates a novel boundary enhancement module for facilitating the interaction between boundary features and spatial features. Finally, to consistently encourage the network to prioritize boundary features, a composite loss function is formulated to guide the network's learning of these features during the decoding stage. Extensive experiments conducted on four ultrasound image datasets demonstrate that GBE-Net surpasses existing state-of-the-art (SOTA) methods.

GBE-Net: Global Boundary Enhancement Network for breast lesion segmentation in ultrasound images