GeoTranMesh: a geometry-guided multi-branch mesh transformer for 3d liver segmentation

Area of Science:

• Medical Imaging
• Computational Geometry
• Surgical Navigation

Background:

• Accurate segmentation of liver surface anatomical landmarks like the falciform ligament and hepatic ridge is critical for intraoperative navigation.
• These landmarks present complex geometry and significant morphological variability, posing challenges for precise identification on 3D liver meshes.

Purpose of the Study:

• To introduce GeoTranMesh, a geometry-guided multi-branch mesh Transformer designed for high-precision segmentation of 3D liver meshes.
• To address the need for accurate segmentation of complex anatomical structures on the liver surface for improved surgical guidance.

Main Methods:

• Employs a hierarchical encoding-decoding structure and global geometric modeling.
• Utilizes a hybrid attention mechanism to fuse local geometric features with cross-branch contextual information.
• Incorporates a directional multi-branch fusion module and geometry-guided multi-task supervision (boundary, distance, normal regression) to enhance geometric consistency and morphological feature learning.

Main Results:

• GeoTranMesh achieved superior segmentation accuracy on a liver mesh dataset.
• Reported Dice scores of 30.9% for the falciform ligament and 66.4% for the liver ridge, with an overall Dice score of 59.5%.
• Demonstrated excellent geometric consistency with a Chamfer distance of 4.4 mm.

Conclusions:

• GeoTranMesh effectively enhances geometric consistency and morphological feature learning through its innovative attention and fusion modules.
• The model achieves precise segmentation of complex anatomical landmarks on the liver.
• Presents significant potential for applications in clinical settings and augmented reality (AR)-guided liver surgery.