MBUSNet: Multimodal breast ultrasound segmentation with attention-guided rectification and boundary enhancement

Breast tumor segmentation in B-mode ultrasound (US) images is challenging due to severe speckle noise, artifacts, lateral shadowing and indistinct boundaries. These inherent characteristics often cause artifacts and low echoes to be misidentified as tumors and result in inaccurate boundaries, making it difficult for single B-mode US image-based methods to achieve reliable segmentation. To address these challenges, a multimodal breast ultrasound image segmentation network (MBUSNet) is proposed that integrates modality-specific edge and spatial information from B-mode US and strain elastography (SE) with boundary-aware augmentation. Firstly, a dual-branch encoder is designed to extract multi-scale feature from both B-mode US and SE images. Subsequently, the attention-guided rectification and aggregation gate (ARAG) is proposed to integrate modality-specific features between multimodal images and achieve feature alignment and fusion. Moreover, a boundary-aware augmented module (BAM) is developed to focus on uncertain boundary region, enhancing the precision of tumor boundary delineation. To evaluate the effectiveness of the proposed method, a multimodal breast ultrasound image dataset was constructed including 604 paired B-mode US and SE images. The proposed MBUSNet achieves superior performance on the dataset with an Intersection over Union (IoU) of 82.67% and a Dice coefficient of 90.05%. Experimental results demonstrate that the proposed method achieves the best overall performance among the compared multimodal methods and provides consistent improvements over comparative single-modal methods.