Pose-aware deep perceptual similarity for iterative 2D/3D registration of knee joints using contrastive learning

Accurate 2D/3D registration between fluoroscopic X-ray images and preoperative CT models is essential to get dynamic joint kinematics in computer-assisted orthopedic analysis. Traditional correlation-based similarity metrics in model-image optimization often suffer from limited capture range and sensitivity to noise and occlusion.

We propose a pose-aware deep perceptual metric using contrastive learning for enlarged capture range and improved robustness. Inspired by LPIPS, our network was specialized and designed for 2D/3D fluoroscopic image registration, where the supervision reflects true 3D transformation discrepancies. The learned metric is integrated into a differentiable 2D/3D registration framework, enabling gradient-based optimization.

The proposed metric enlarges the capture range of 2D/3D registration and smooths the loss landscape, while ensuring the required accuracy for automated registration. Our metric outperforms conventional similarity measures and standard LPIPS, presenting zero-shot ability for unseen low-contrast fluoroscopic conditions.

The proposed metric provides a robust and differentiable metric for fluoroscopic 2D/3D registration, improving accuracy in joint kinematics analysis for both single- and dual-plane fluoroscopy.