Uncertainty Quantification in Image-based 2D/3D Registration and Its Relationship with Accuracy.

Area of Science:

• Medical Imaging
• Computer Vision
• Robotics

Background:

• Accurate 2D/3D registration is vital for image-guided navigation and surgical robotics.
• Estimating and interpreting uncertainty in 2D/3D registration is challenging due to dimensional inconsistencies.
• Existing methods struggle with reliable uncertainty quantification in this domain.

Purpose of the Study:

• To develop and characterize a novel method for uncertainty quantification in single-view 2D/3D registration.
• To address the specific challenges of estimating uncertainty in 2D/3D registration tasks.
• To investigate the relationship between quantified uncertainty and actual registration error.

Main Methods:

• Modeled 2D/3D registration as a Maximum A Posteriori (MAP) estimation.
• Quantified uncertainty by sampling from an approximate posterior distribution.
• Generated synthetic 2D/3D pelvis registrations for experimental validation.

Main Results:

• XGBoost regression demonstrated a strong fit (R-squared = 0.85) for uncertainty-registration error relationship, outperforming OLS (R-squared = 0.023).
• Significant differences in prediction accuracy were observed across registration error groups.
• Uncertainty metrics showed differing importance depending on the model's focus (global vs. low-error regimes).

Conclusions:

• Presented a novel approach for uncertainty quantification and characterization in single-view 2D/3D registration.
• Revealed a nonlinear correlation between uncertainty and registration accuracy, particularly in low-error scenarios.
• Provided foundational insights for enhancing the reliability of image-guided interventions through better uncertainty understanding.