HTBE-Net: A hybrid transformer network based on boundary enhancement for breast ultrasound image segmentation

Automatic segmentation algorithms for breast ultrasound images are crucial for early breast cancer detection and treatment. Existing methods centered on Convolutional Neural Networks (CNNs) and Transformers has made great strides by focusing on the development of multi-branch coding networks with multiple receptive fields. However, there are still challenges in the practical application of these methods. Current methods tend to coarsely fuse features from each branch, and the lack of effective feature interaction between high-dimensional features with different receptive fields leads to models that do not take full advantage of the diverse perspectives of the breast lesion regions. In addition, the coarse-grained feature interaction strategy tends to lead to the blurring of lesion boundaries. To address the above challenges, a novel dual-branch automatic segmentation algorithm, named HTBE-Net, is proposed in this paper. Specifically, a Boundary Guided Module (BGM) is firstly designed to guide the encoder to outline the precise lesion regions. As a complement, a Selective Feature Enhancement Module (SFEM) is designed and applied to each branch of the encoder to highlight the weights of the boundary features. Finally, a Long-Short Range Attention Interaction Fusion (LSIF) module was designed to carefully fuse encoder features from different branches. This module facilitates the feature interaction between the features and utilizes different receptive fields to optimize the network's segmentation of the lesion regions. Extensive experiments based on three ultrasound image datasets show that HTBE-Net outperforms existing state-of-the-art (SOTA) methods.