A boundary-semantic synergy learning network for breast ultrasound lesion segmentation
TL;DR
Results indicate that D-SAS provides accurate breast lesion segmentation with promising cross-dataset generalization capability and strong robustness to domain shifts caused by different imaging devices and clinical centers.
Abstract
Breast ultrasound (BUS) image segmentation remains challenging due to speckle noise, blurred lesion boundaries, and substantial variations in lesion scale. To address these issues, we propose a Dual-branch Semantic-Aware Segmentation Network (D-SAS) that enhances lesion representation through boundary enhancement, multi-scale semantic perception, and boundary-semantic collaborative fusion. Specifically, a Speckle Suppression Boundary Gate (SSBG) is designed to suppress noise interference and strengthen boundary representation. An Adaptive Atrous Selector (AAS) is introduced to dynamically aggregate multi-scale contextual information via pixel-wise adaptive weighting. In addition, a Synergistic Fusion Head (SFH) facilitates effective interaction between boundary and semantic features through bidirectional attention guidance and uncertainty-aware fusion. Experimental results demonstrate the effectiveness of the proposed method. D-SAS achieves Dice scores of 82.10% and 89.14% and IoU values of 70.21% and 81.34% on the BUSI and BUSI_WHU datasets, respectively, outperforming existing state-of-the-art methods. Furthermore, under the cross-dataset setting of BUSI_WHU-to-UDIAT, D-SAS achieves a Dice score of 80.95% and an IoU of 72.82%, demonstrating strong robustness to domain shifts caused by different imaging devices and clinical centers. These results indicate that D-SAS provides accurate breast lesion segmentation with promising cross-dataset generalization capability.