Lesion segmentation method of breast ultrasound images based on residual attention U-Net and fusion dataset

Abstract

Breast ultrasound is a crucial imaging modality for early breast cancer screening and benign-malignant lesion differentiation. Accurate lesion segmentation serves as an essential prerequisite for clinical quantitative evaluation and intelligent diagnosis and treatment. The residual attention fused U-Net is constructed in this paper to address the limitations of conventional U-Net, which possesses insufficient deep feature representation capability, fails to adapt to blurred lesion boundaries and complex backgrounds of breast ultrasound images, and suffers from poor edge segmentation performance and inadequate model generalization. Relying on residual structures, the proposed model strengthens multi-scale feature learning, and adopts attention gates to concentrate lesion features and suppress irrelevant background interference. Cross-domain datasets including BUSI and BUS-UCLM are fused, and multiple image preprocessing and augmentation methods are adopted to optimize data distribution. Experimental results demonstrate that compared with the original U-Net, the Dice coefficient of the proposed model increases from 0.6608 to 0.8737, and the IoU rises from 0.5914 to 0.8258. The segmentation accuracy and model robustness are significantly improved, which can accurately restore lesion contours and provide reliable technical support for intelligent auxiliary diagnosis of breast ultrasound.