How can I improve 3D liver imaging for surgery?

Abstract

<h4>Purpose</h4>Anatomical landmarks on the liver surface, such as the falciform ligament and hepatic ridge, exhibit complex geometry and significant morphological variability. Accurate segmentation of these structures on 3D liver mesh is essential for intraoperative navigation. This paper presents GeoTranMesh, a geometry-guided multi-branch mesh Transformer that employs hierarchical encoding-decoding and global geometric modeling to achieve high-precision liver mesh segmentation.<h4>Methods</h4>A hybrid attention mechanism is proposed to fuse local geometric features with cross-branch contextual information, while a directional multi-branch fusion module refines features along tangent, normal, and bitangent directions to enhance geometric consistency. In addition, geometry-guided multi-task supervision, including boundary, distance, and normal regression, is incorporated to strengthen morphological feature learning.<h4>Results</h4>On liver mesh dataset, GeoTranMesh achieved the highest segmentation accuracy, with Dice scores of 30.9% and 66.4% for the falciform ligament and liver ridge, respectively, an overall Dice of 59.5%, and a Chamfer distance of only 4.4 mm, demonstrating superior geometric consistency and anatomical precision.<h4>Conclusion</h4>GeoTranMesh integrates hybrid attention and directional multi-branch fusion to enhance geometric consistency and morphological feature learning, achieving precise segmentation of complex anatomical landmarks, demonstrating potential for clinical and AR-guided liver surgery applications.