Neural Implicit Shape and Intensity Models for Scan-Free 2D-3D Registration in Dynamic Stereo-Radiography

2D-3D registration in dynamic stereo-radiography is performed in biomechanics and orthopedics methods to acquire 6-degree of freedom poses of native anatomy. Conventional registration frameworks rely on subject-specific digitized anatomy segmented from computed tomography scans, which increases processing efforts associated with registration, and also presents additional radiation risks. Scan-free registration has thus been introduced to circumvent reliance on volumetric imaging by jointly estimating shape and pose in radiographs. Existing approaches utilize principal component analysis for efficient sampling of plausible anatomic candidates, but this technique involves tedious and error-prone mesh or scan registration as a developmental step. The current work proposes an alternative method for scan-free 2D-3D registration, referred to as Neural Implicit Shape and Intensity Models.

Neural Implicit Shape and Intensity Models were developed to capture population-level anatomic shape and intensity variability observed across a training set, without reliance on mesh registration. Trained models were integrated with a 2D-3D registration framework for pose estimation and anatomy reconstruction in stereo-radiographs.

The framework was evaluated by performing 2D-3D registration of the native knee in stereo-radiographs capturing in vivo movement, and in additional frames capturing a synthetic leg phantom. Experiments revealed geometric errors consistently near or below 1 mm, and pose errors near or below 1 $_{}^{\circ }$ or mm in both evaluation cohorts.

This work proposed novel methods for scan-free 2D-3D registration in stereo-radiography. Results indicate the introduced framework may benefit registration applications by addressing dependence on volumetric medical imaging.