MSFA-Net: Multiscale feature enhancement and fusion attention-based dual-encoder network for breast ultrasound segmentation

Precise segmentation of breast ultrasound images is vital for the early diagnosis of breast cancer and remains a challenging task. Existing segmentation methods based on convolutional neural networks often have limited receptive fields, leading to inaccurate segmentation of blurred boundaries and irregular lesion shapes in breast ultrasound images. Therefore, we propose a multiscale feature enhancement and fusion attention-based dual-encoder network. Our design is as follows: Firstly, we introduce transformer as a global context-guided encoding branch to establish long-term dependencies. Secondly, with fewer parameters, we propose an efficient auxiliary encoder, which introduces multiscale feature enhancement module and fusion attention module. This design facilitates feature interaction across receptive fields, alleviates the gridding problem, and enhances the model's ability to capture fine-grained local and global features. Thirdly, full-stage adaptive cross-attention fusion dynamically adjust the focus areas and scales, thereby effectively integrating multi-layer information. Extensive experiments are conducted on three publicly available ultrasound datasets, comparing it with 12 state-of-the-art methods. Notably, our model outperforms the suboptimal method, improving Dice metric by 2.56% on the BUS dataset and 2.19% on the BUSI-malignant dataset.