Scholarlink

<p>Breast ultrasound image segmentation is essential for accurate diagnosis, yet ultrasound images often have low resolution and poor signal-to-noise ratio, making abnormal lesions difficult to distinguish. Manual segmentation by doctors is also labor-intensive and time-consuming. Although the Segment Anything Model (SAM) shows strong zero-shot segmentation on natural images, it performs poorly on ultrasound images due to blurred boundaries, speckles, and diverse contrast. To address this, we propose BUSI-SAM, which fine-tunes SAM with breast ultrasound datasets and SA-1B pretrained weights. Experiments on the BUSI and BUSI-WHU datasets show that BUSI-SAM achieves superior performance, with mIoU of 89.29%, OA of 98.59%, and HD95 of 11.12 on BUSI, and mIoU of 90.96%, OA of 98.86%, and HD95 of 10.37 on BUSI-WHU. Precision-Recall and ROC curves also outperform other algorithms. BUSI-SAM demonstrates robustness to noise, shadows, and contrast variations, enabling reliable lesion boundary delineation even in challenging conditions. Its high accuracy and generalization make it a valuable clinical tool to reduce manual workload, support tumor boundary annotation, and enhance early breast cancer diagnosis.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Segment Anything for Breast Cancer Ultrasound Image

  • 期刊论文

  • 发表日期：

    2025年12月13日

查看全文
  • Jin Huang,
  • Yimin Zhang,
  • Yujie Xie,
  • Jingwen Deng,
  • Du Wang,
  • Liye Mei,
  • Cheng Lei

摘要

Breast ultrasound image segmentation is essential for accurate diagnosis, yet ultrasound images often have low resolution and poor signal-to-noise ratio, making abnormal lesions difficult to distinguish. Manual segmentation by doctors is also labor-intensive and time-consuming. Although the Segment Anything Model (SAM) shows strong zero-shot segmentation on natural images, it performs poorly on ultrasound images due to blurred boundaries, speckles, and diverse contrast. To address this, we propose BUSI-SAM, which fine-tunes SAM with breast ultrasound datasets and SA-1B pretrained weights. Experiments on the BUSI and BUSI-WHU datasets show that BUSI-SAM achieves superior performance, with mIoU of 89.29%, OA of 98.59%, and HD95 of 11.12 on BUSI, and mIoU of 90.96%, OA of 98.86%, and HD95 of 10.37 on BUSI-WHU. Precision-Recall and ROC curves also outperform other algorithms. BUSI-SAM demonstrates robustness to noise, shadows, and contrast variations, enabling reliable lesion boundary delineation even in challenging conditions. Its high accuracy and generalization make it a valuable clinical tool to reduce manual workload, support tumor boundary annotation, and enhance early breast cancer diagnosis.