A personalized 2D/3D image registration pipeline for automated knee registration using single-plane moving fluoroscopy

Abstract Accurate knee joint kinematics are essential for biomechanical research and clinical assessment. Single-plane fluoroscopy is one of the most accurate tools for measuring in vivo joint motion, but manual estimation of three-dimensional bone pose from individual fluoroscopic frames remains labor-intensive and time-consuming. In this study, we present a personalized 2D/3D image registration pipeline for automated registration of knee fluoroscopic images acquired using single-plane moving fluoroscopy. The proposed framework combines pose standardization across acquisition sessions, subject-specific learning-based pose estimation, and differentiable-rendering-based pose refinement. Tested using fluoroscopic datasets of both natural knees and knee implants together with a PyTorch3D-based differentiable renderer, we trained subject-specific pose estimation networks for both anatomy types. Evaluation against manually matched reference poses showed mean initial in-plane translation differences of approximately 1.1–2.1 mm and geodesic rotation differences of 1.3–3.6 $$^\circ$$ with mask-only input, while out-of-plane translation differences remained larger, as expected in single-plane imaging. Gradient-based refinement using gradient correlation coefficient (GCC) further reduced the disagreement with the manual reference poses, resulting in sub-millimetre in-plane translation differences and sub-degree rotational differences, with over 95% of test frames satisfying the predefined success criterion that both the femoral and tibial registrations had absolute in-plane translation differences below 1 mm in both $$T_x$$ and $$T_y$$ and a geodesic rotation difference below 1 $$^\circ$$ . These results quantify agreement with the existing manual registration workflow rather than independently validated anatomical pose accuracy.