Toward Patient-Specific Partial Point Cloud to Surface Completion for Pre to Intra-operative Registration in...

Area of Science:

• Medical Imaging
• Computer-Aided Surgery
• Geometric Deep Learning

Background:

• Intra-operative data in image-guided surgery lacks sub-surface details crucial for identifying vessels and tumors.
• Image-to-physical registration, vital for fusing pre-operative and intra-operative data, is hindered by incomplete intra-operative point clouds.

Purpose of the Study:

• To develop a patient-specific point cloud completion method to enhance image-to-physical registration in image-guided surgery.
• To evaluate the effectiveness of VN-OccNet in generating complete liver surfaces from partial intra-operative data.

Main Methods:

• Leveraged VN-OccNet for patient-specific point cloud completion, training with simulated deformations.
• Analyzed VN-OccNet's rotation-equivariance for surface recovery from partial data.
• Integrated completed surfaces into the Go-ICP algorithm for rigid registration improvement.

Main Results:

• Demonstrated improved initial rigid registration outcomes using the completed intra-operative surface.
• Showcased VN-OccNet's capability in recovering complete surfaces from partial point clouds.
• Validated the utility of patient-specific completion for addressing partial visibility challenges.

Conclusions:

• Patient-specific point cloud completion shows promise for robust registration in image-guided surgery.
• VN-OccNet's rotation-equivariant and surface generation abilities are valuable for registration frameworks.
• The approach mitigates challenges associated with partial intra-operative visibility.