Couinaud segment-aware deep learning on point clouds for major liver resection planning.

Area of Science:

• Medical Imaging
• Surgical Planning
• Deep Learning

Background:

• Automatic liver resection planning is crucial for major surgeries like hemi-hepatectomy.
• Couinaud liver segments are vital for anatomical description and surgical decision-making in liver procedures.
• Current deep learning models may not fully leverage this anatomical information.

Purpose of the Study:

• To investigate the impact of incorporating Couinaud liver segment information on deep learning model performance for automatic liver resection planning.
• To enhance the clinical relevance and accuracy of AI-driven surgical planning for major liver resections.

Main Methods:

• A point cloud-based geometric deep learning approach using a modified RandLA-Net architecture was developed.
• The model was trained and validated on internal (70 cases) and external (30 cases) hemi-hepatectomy datasets.
• Two composite loss functions (CE+IoU and CE+Dice) were evaluated with and without Couinaud segment data.

Main Results:

• The CE+IoU loss function generally outperformed CE+Dice.
• Incorporating Couinaud segment information significantly improved mean IoU (e.g., from 0.787 to 0.804 on the internal set) and F1-score.
• Statistical analysis confirmed significant performance gains, particularly in preserving critical vascular structures.

Conclusions:

• Explicitly integrating Couinaud segment information enhances quantitative performance in automatic liver resection planning.
• This approach improves clinical relevance by better preserving vital anatomical structures during simulated resections.
• Deep learning models incorporating anatomical data offer a promising tool for improving surgical outcomes in liver surgery.