Scholarlink

<p>Medical image segmentation is critical for assisted diagnosis and treatment evaluation, yet existing approaches still struggle to generalize across diverse datasets, often relying on manual prompts and exhibiting insufficient boundary and fine-grained alignment. We propose HeterSAM, which performs efficient alignment between prompt and image features via heterogeneous representation alignment and deformable bidirectional cross-attention (DCA), thereby strengthening boundary and detail modeling and improving accuracy without manual interaction. We introduce learnable prompt tokens as global structural priors to enhance target specificity and cross-domain stability. By combining a dense prompt encoder with a DCA-based decoder, HeterSAM achieves state-of-the-art results on multiple benchmarks, including skin lesions, colorectal polyps, and spinal structures. Experiments show significant gains in IoU and cross-dataset transferability, indicating robust generalization. In addition, we provide a lightweight decoder as a deployment alternative that reduces inference overhead with minimal accuracy loss. Our code is open-sourced at: <a href="https://github.com/adgnahs/HeterSAM.git">https://github.com/adgnahs/HeterSAM.git</a>.</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

Enhancing medical image segmentation with heterogeneous representation alignment and deformable attention

  • 期刊论文

  • 发表日期：

    2026年4月14日

查看全文
  • Da Shang,
  • Kainan Leu,
  • Ying Yu,
  • Zihao Jia,
  • Yifan Li,
  • Henghui Meng

摘要

Medical image segmentation is critical for assisted diagnosis and treatment evaluation, yet existing approaches still struggle to generalize across diverse datasets, often relying on manual prompts and exhibiting insufficient boundary and fine-grained alignment. We propose HeterSAM, which performs efficient alignment between prompt and image features via heterogeneous representation alignment and deformable bidirectional cross-attention (DCA), thereby strengthening boundary and detail modeling and improving accuracy without manual interaction. We introduce learnable prompt tokens as global structural priors to enhance target specificity and cross-domain stability. By combining a dense prompt encoder with a DCA-based decoder, HeterSAM achieves state-of-the-art results on multiple benchmarks, including skin lesions, colorectal polyps, and spinal structures. Experiments show significant gains in IoU and cross-dataset transferability, indicating robust generalization. In addition, we provide a lightweight decoder as a deployment alternative that reduces inference overhead with minimal accuracy loss. Our code is open-sourced at: https://github.com/adgnahs/HeterSAM.git.