To develop a deep perceptual similarity metric that is aware of pose variations to improve the accuracy of 2D/3D registration in fluoroscopic imaging.
Approach:
Key Findings:
The proposed metric yields smooth and robust loss landscapes, improving convergence reliability in both single-plane and dual-plane fluoroscopic scenarios.
The dynamic pose-dependent margin stabilizes training across large perturbation ranges, enhancing capture range while preserving convexity.
Interpretation:
The developed metric aims to address limitations of existing similarity measures by providing a convex similarity landscape that is intended to generalize across rigid pose variations, particularly under noisy imaging conditions.
Limitations:
Existing perceptual similarity networks are typically trained on natural images, which may not reflect the specific challenges of fluoroscopic imaging, such as radiographic physics and occlusions.
The generalization of the proposed method to other registration problems may require further validation and testing in diverse imaging scenarios.
Conclusion:
The study presents a novel approach to enhance 2D/3D registration accuracy in knee joint analysis through a pose-aware deep perceptual similarity metric.
