Intraoperative estimation of liver boundary conditions from multiple partial surfaces

Area of Science:

• Medical Imaging
• Surgical Robotics
• Computational Anatomy

Background:

• Accurate intraoperative anatomical information is crucial for computer-assisted surgery.
• Patient-specific biomechanical models (PBMs) are vital for predicting tissue deformations.
• Estimating anatomical attachments preoperatively is challenging and impacts PBM accuracy.

Purpose of the Study:

• To develop a deep neural network for predicting the location of tissue attachments.
• To leverage intraoperative deformed organ surface data for improved attachment prediction.
• To enhance the accuracy of patient-specific biomechanical models in computer-assisted surgery.

Main Methods:

• A deep neural network was trained using multiple partial views of intraoperative deformed organ surfaces encoded as point clouds.
• The network utilized a sequence of deformed views to capture the temporal evolution of tissue deformations.
• This approach addressed the ambiguity inherent in estimating attachments from single-view data.

Main Results:

• The method was applied to computer-assisted hepatic surgery, validated on synthetic and in vivo human data.
• The network achieved a 26% improvement in prediction accuracy compared to existing methods.
• Training on patient-specific synthetic data was completed in under 5 hours, demonstrating rapid deployment.

Conclusions:

• The proposed deep neural network accurately and rapidly estimates tissue attachments by considering deformation dynamics.
• Patient-specific simulated data enables effective training for real-time surgical guidance.
• This advancement significantly improves intraoperative anatomical guidance in computer-assisted surgical systems.