An Unsupervised Fluoroscopic Analysis of Knee Joint Kinematics

Knowledge of the three dimensional positions of bones at a joint as a function of time is required to accurately model joint kinematics. 3-D bone geometry data from a static computer tomography (CT) images can be combined with time sequence information from 2-D video fluoroscopy images to produce 3-D position data over time. The process involves creating virtual X-rays from the CT image through digitally reconstructed radiograph (DRR) projections. Historically, the process of matching the 3-D and 2-D data has required human interaction. We have eliminated the need for manual initialization using a Monte Carlo technique with a variable search range. The search range decreases as the matching improves, avoiding the inherent problems of local minima in the optimization search. Experiments demonstrate that image positions can be matched to within 1 degree rotation, azimuth and elevation without human intervention