Scholarlink

<p>While the Segment Anything Model (SAM) excels in natural image segmentation, its application to medical imaging, especially ultrasound, faces challenges like artifacts and low-contrast boundaries, hindering domain adaptation. Existing methods often underutilize SAM’s pre-trained potential, inadequately optimizing both parameter and feature spaces, leading to suboptimal trade-offs between efficiency and accuracy. This paper introduces HATSAM (Hierarchical Adaptation Strategy for Segment Anything Model), a hierarchical adaptation network for efficient SAM transfer in medical image segmentation. HATSAM addresses parameter space adaptability in the image encoder with a Parameter Space Completeness Dynamic Adaptation (PSCDA) module, which combines adapter-based fine-tuning for incomplete parameters with SVD-based singular value updates for relatively complete ones. For feature space misalignment in the mask decoder, the Detail Enhanced Feature Fusion Module (DEFFM) denoises fine-grained features and integrates them with global ones. HATSAM effectively adapts SAM for ultrasound segmentation with lower computational cost. Extensive experiments on seven public datasets (TN3K, TG3K, BUSI, CAMUS, DDTI, UDIAT, EchoCP) show HATSAM achieves a 21.32% inference speed improvement and a 36.97% computational cost reduction over state-of-the-art approaches, while maintaining competitive accuracy.</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

HATSAM: hierarchical adaptation strategy for segment anything model in medical imaging

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Tinglong Tang,
  • Zhe Rao,
  • Yuan Wang,
  • Sixian Chan,
  • Chao Li,
  • Ye Sang,
  • Lili Zou

摘要

While the Segment Anything Model (SAM) excels in natural image segmentation, its application to medical imaging, especially ultrasound, faces challenges like artifacts and low-contrast boundaries, hindering domain adaptation. Existing methods often underutilize SAM’s pre-trained potential, inadequately optimizing both parameter and feature spaces, leading to suboptimal trade-offs between efficiency and accuracy. This paper introduces HATSAM (Hierarchical Adaptation Strategy for Segment Anything Model), a hierarchical adaptation network for efficient SAM transfer in medical image segmentation. HATSAM addresses parameter space adaptability in the image encoder with a Parameter Space Completeness Dynamic Adaptation (PSCDA) module, which combines adapter-based fine-tuning for incomplete parameters with SVD-based singular value updates for relatively complete ones. For feature space misalignment in the mask decoder, the Detail Enhanced Feature Fusion Module (DEFFM) denoises fine-grained features and integrates them with global ones. HATSAM effectively adapts SAM for ultrasound segmentation with lower computational cost. Extensive experiments on seven public datasets (TN3K, TG3K, BUSI, CAMUS, DDTI, UDIAT, EchoCP) show HATSAM achieves a 21.32% inference speed improvement and a 36.97% computational cost reduction over state-of-the-art approaches, while maintaining competitive accuracy.