乳腺癌是全球女性中最常见的恶性肿瘤之一，准确的病变分割对于乳腺癌的早期诊断与治疗具有重要意义。然而，由于病变形态的多样性以及超声成像机制的复杂性，现有基于深度学习的乳腺超声图像病变分割方法在分割准确性方面仍面临巨大挑战。为进一步提升乳腺超声图像中病变区域的分割精度，基于经典U-Net架构提出了一种新型乳腺超声图像病变分割网络CWSASK-BBA-Net。首先，在网络中引入逐通道空间自适应选择核卷积模块（CWSASKM），根据不同通道的语义特征为每个空间位置自适应选择感受野大小，从而增强多尺度信息的建模能力。其次，引入双向边界感知机制（BBAM），通过联合正向与反向注意力协同建模目标显著区域及其边界，逐步提升对非显著区域的判别能力，进一步强化边界信息的表达。最后，为验证CWSASK-BBA-Net的性能，本文在三组公开乳腺超声图像数据集上开展了分割实验。结果显示，该方法在BUSI数据集上的杰卡德指数、精确率、召回率和Dice相似系数分别为71.97%、82.85%、81.40%和80.44%，较次优方法分别提升1.69%、1.05%、1.28%和1.84%；在UDIAT数据集上，四项指标分别达到78.14%、88.31%、86.73%和86.10%，对应提升2.75%、2.04%、0.56%和2.01%；在外部数据集STU上也取得了优于其他方法的整体表现。实验结果表明，CWSASK-BBA-Net在乳腺超声图像分割任务中展现出更优的整体性能。

Breast cancer is one of the most common malignant tumors among women worldwide. Accurate lesion segmentation plays a crucial role in the early diagnosis and treatment of breast cancer. However, due to the variability of lesion morphology and the complexity of ultrasound imaging mechanisms, existing deep learning-based methods for breast ultrasound lesion segmentation still face significant challenges in segmentation accuracy.To further improve the segmentation precision of lesion regions in breast ultrasound images, this paper proposes a novel lesion segmentation network, CWSASK-BBA-Net, based on the classic U-Net architecture. First, a Channel-wise Spatially Adaptive Selective Kernel Convolution Module (CWSASKM) is introduced, which adaptively selects the receptive field size at each spatial location based on the semantic features of different channels, thereby enhancing the network’s capability to model multi-scale information. Second, a Bidirectional Boundary-Aware Mechanism (BBAM) is integrated to jointly model salient targets and their boundaries through forward and reverse attention, progressively improving the discrimination of non-salient regions and further strengthening the representation of boundary information.To validate the performance of CWSASK-BBA-Net, segmentation experiments were conducted on three publicly available breast ultrasound image datasets. Results show that the proposed method achieves average values of 71.97%, 82.85%, 81.40%, and 80.44% on the BUSI dataset in terms of Jaccard index, precision, recall, and Dice similarity coefficient, respectively, with improvements of 1.69%, 1.05%, 1.28%, and 1.84% compared to the second-best methods. On the UDIAT dataset, the four metrics reach 78.14%, 88.31%, 86.73%, and 86.10%, with gains of 2.75%, 2.04%, 0.56%, and 2.01%, respectively. In addition, the method also demonstrates overall performance advantages on the external STU dataset. These results indicate that CWSASK-BBA-Net exhibits superior overall performance in breast ultrasound image segmentation tasks.