Robust and smooth Couinaud segmentation via anatomical structure-guided point-voxel network

Area of Science:

• Medical imaging analysis
• Computational anatomy
• Surgical planning

Background:

• Precise Couinaud segmentation of liver segments from computed tomography (CT) is vital for surgical planning and lesion analysis.
• Segmentation is difficult due to reliance on vessel structures and lack of intensity contrast between adjacent liver segments in CT scans.

Purpose of the Study:

• To develop a novel framework for robust and smooth Couinaud segmentation of liver segments from CT images.
• To address the challenges of vessel-based segmentation and inter-segment contrast limitations.

Main Methods:

• A multi-scale point-voxel fusion framework was designed, processing 3D liver point clouds and voxel grids with embedded vessel structures.
• The framework utilizes two input-specific branches for complementary feature extraction from points and voxels.
• A local attention module adaptively fuses features across scales, and a novel distance loss enhances inter-segment feature compactness.

Main Results:

• The proposed method demonstrated superior performance on three public liver datasets compared to state-of-the-art techniques.
• Outperformed 3D UNet by ~20% and PointNet2Plus by ~8% in Dice score on the LiTS dataset's out-of-distribution testing.
• Achieved robust and smooth Couinaud segmentations, improving segmentation certainty.

Conclusions:

• The multi-scale point-voxel fusion framework effectively models spatial relationships and semantic information for improved liver segmentation.
• The method offers a significant advancement in automated Couinaud segmentation, aiding surgical planning and liver lesion examination.
• Code and annotations are publicly available to facilitate further research.