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

Area of Science:

• Medical Imaging
• Computer-Aided Surgery
• Geometric Deep Learning

Background:

• 3D rigid registration is essential for augmented reality (AR) navigation in laparoscopic liver surgery (LLS).
• Challenges include smooth liver surfaces, density variations, and partial visibility, hindering accurate registration.
• Existing methods struggle with these specific challenges in LLS.

Purpose of the Study:

• To develop an advanced 3D rigid registration technique for LLS.
• To enhance feature discrimination and registration robustness using geometric information.
• To improve the accuracy of augmented reality navigation in liver surgery.

Main Methods:

• Proposed a geometry-informed attention network (GIA-Net) for 3D rigid registration.
• GIA-Net integrates surface curvature and normal vectors via a Geometry Transformer.
• A Representative Point Selector balances point cloud density for improved accuracy.

Main Results:

• GIA-Net demonstrated lower transformation errors compared to existing methods on multiple datasets (MedShapeNet, 3Dircadb, DePoLL).
• Achieved a 2.23 cm reduction in mean target registration error (TRE) on the DePoLL dataset.
• Results indicate superior performance in challenging LLS scenarios.

Conclusions:

• Geometric information significantly boosts rigid registration performance in LLS.
• GIA-Net provides a reliable foundation for nonrigid refinement in surgical navigation.
• The proposed method enhances the feasibility of AR-guided liver surgery.