Scholarlink

Purpose <p>3D rigid registration in laparoscopic liver surgery (LLS) is crucial for augmented reality navigation. However, liver surface registration faces several challenges, including the smooth surface with few distinctive textures, density variations between preoperative and intraoperative point clouds, and partial visibility of intraoperative surfaces.</p> Methods <p>To address these issues, we propose a geometry-informed attention network (GIA-Net) that leverages surface curvature and normal vectors to improve feature discrimination. GIA-Net includes two key components: a Geometry Transformer that integrates geometric priors for more distinctive feature representation, and a Representative Point Selector that balances local density between point clouds to enhance registration robustness and accuracy.</p> Results <p>Experiments on the MedShapeNet, 3Dircadb, and DePoLL datasets show that GIA-Net achieves lower transformation errors than existing rigid registration methods. On the DePoLL dataset, our method reduces the mean target registration error (TRE) by 2.23 cm compared to the baseline.</p> Conclusions <p>Leveraging geometric information significantly enhances rigid registration performance, providing a reliable foundation for subsequent nonrigid refinement in liver surgery navigation.</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

Gia-Net: geometry-informed attention network for 3D point cloud registration in liver surgery

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • Xiaoyue Liu,
  • Tian Xu,
  • Ziyi Jin,
  • Xiang Deng,
  • Tianyuan Gan,
  • Zuoming Fu,
  • Chongan Zhang,
  • Peng Wang,
  • Hong Zhang,
  • Liangliang Yu,
  • Xiao Liang,
  • Liping Qin,
  • Xuesong Ye

摘要

Purpose

3D rigid registration in laparoscopic liver surgery (LLS) is crucial for augmented reality navigation. However, liver surface registration faces several challenges, including the smooth surface with few distinctive textures, density variations between preoperative and intraoperative point clouds, and partial visibility of intraoperative surfaces.

Methods

To address these issues, we propose a geometry-informed attention network (GIA-Net) that leverages surface curvature and normal vectors to improve feature discrimination. GIA-Net includes two key components: a Geometry Transformer that integrates geometric priors for more distinctive feature representation, and a Representative Point Selector that balances local density between point clouds to enhance registration robustness and accuracy.

Results

Experiments on the MedShapeNet, 3Dircadb, and DePoLL datasets show that GIA-Net achieves lower transformation errors than existing rigid registration methods. On the DePoLL dataset, our method reduces the mean target registration error (TRE) by 2.23 cm compared to the baseline.

Conclusions

Leveraging geometric information significantly enhances rigid registration performance, providing a reliable foundation for subsequent nonrigid refinement in liver surgery navigation.