Assessing Learning-Based Reconstructed Liver Surfaces From Partial Point Clouds for Improving Pre- to Intra-Operative...

Area of Science:

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

Background:

• Image-guided liver surgery requires accurate fusion of pre-operative and intra-operative data.
• Partial visibility of the liver during surgery poses a significant challenge for registration.
• Learning-based methods for generating complete surfaces from partial point clouds offer potential solutions.

Purpose of the Study:

• To analyze the error introduced by generating complete intra-operative liver surfaces from partial data.
• To evaluate the impact of this generated surface error on both rigid and non-rigid registration algorithms.
• To ensure robust performance of image-guided liver surgery in clinical settings.

Main Methods:

• Utilized a VN-OccNet framework trained on simulated deformed liver data.
• Generated complete surfaces from partial point clouds acquired from multiple viewpoints of in vitro liver phantoms.
• Integrated generated surfaces into Go-ICP and GMM-FEM registration frameworks.
• Estimated registration errors in both visible and invisible regions.

Main Results:

• Surface generation error increases with distance from the visible partial surface.
• Generated surface errors impact registration accuracy in invisible regions.
• Registration errors were also observed in visible regions, affecting accuracy within the camera's field of view.

Conclusions:

• Understanding and quantifying surface generation error is crucial for accurate image-guided liver surgery.
• The proposed method provides insights into error propagation during registration with generated surfaces.
• Further refinement is needed to mitigate errors in both visible and invisible regions for clinical application.