Learning an Attention Model for Robust 2-D/3-D Registration Using Point-To-Plane Correspondences.

Area of Science:

• Medical Imaging
• Computer-Aided Surgery
• Machine Learning

Background:

• Minimally invasive procedures require precise image guidance for navigation.
• X-ray images in surgery often lack visibility of critical anatomical structures.
• Accurate 2-D/3-D registration is essential for overlaying pre-operative 3-D data onto intra-operative 2-D images.

Purpose of the Study:

• To develop a learning-based correspondence weighting scheme to improve 2-D/3-D image registration robustness.
• To enhance the accuracy and reliability of image guidance in minimally invasive interventions.

Main Methods:

• A novel attention model was learned to weight correspondences, prioritizing inliers and suppressing outliers.
• The model was trained using an objective function that minimizes registration error, avoiding per-correspondence labels.
• The method was evaluated on its ability to generalize across different views and anatomical structures.

Main Results:

• The learning-based weighting scheme significantly increased registration success rates, e.g., from 84.9% to 97.0% for spine registration.
• Achieved high accuracy with a mean re-projection distance of approximately 0.5mm.
• Demonstrated robustness and generalization capabilities, requiring minimal training data and learning from simulated data.

Conclusions:

• The proposed learning-based correspondence weighting method substantially improves the robustness and accuracy of 2-D/3-D image registration.
• This technique offers a more reliable image guidance solution for minimally invasive surgery.
• The method's ability to generalize and learn from limited data makes it broadly applicable in surgical navigation.