Dual-branch adaptive fusion network with edge supervision for breast ultrasound image segmentation

Breast ultrasound image segmentation is a crucial task in computer-aided diagnosis , enabling rapid localization of the studied lesion regions. However, for breast ultrasound image with small, blurry, and irregularly shaped lesions, many current methods solely rely on supervised segmentation based on masks, neglecting the relevant edge information. In this article, we propose a simple and direct edge supervision (ES) block to pay additional attention to the edge information of breast lesions to improve the edge quality in the segmentation results. The proposed block can be applied not only at the final layer of the network but also at intermediate layers to enhance supervision of the edge in the segmented regions, thereby improving the segmentation quality. We also propose an adap-tive feature fusion (AFF) block to efficiently combine the dual-branch features of local and global information. Specifically, we construct a dual-branch adaptive fusion network based on CNNs and Transformers to learn the feature representation of the breast ultrasound image. Additionally, we introduce the proposed edge supervision (ES) block, which employs a hierarchical supervision approach to learn edge features. Our method achieves remarkable segmentation results on two widely adopted breast ultrasound image datasets, BUSI and UDIAT. Extensive ablation experiments confirm the effectiveness of our method in the fusion of CNNs and Transformers, as well as the edge supervision. Comparison with several existing methods demonstrates that the proposed approach achieves competitive performance in breast ultrasound image segmentation. And robustness experiment demonstrates the high generalization capability of our method. (a) benign (b) malignant Fig. 1 Breast ultrasound images in the area of the lesion, edge irregularities, and similarities in the edge and surrounding areas.