Abstract
The standard approach for kinematic analysis of knee joints has been roentgen stereophotogrammetry (RSA). This approach requires implanting tantalum beads during surgery so pre- and post-surgery comparisons have not been conducted. CT- fluoroscopy registration is a non-invasive alternative but has had accuracy and speed limitations. Our new algorithm addresses these limitations.
Our approach to the problem of registering CT data to single-plane fluoroscopy was to generate a digitally reconstructed radiograph (DRR) from the CT data and then filter this to produce an edge-enhanced image, which was then registered with an edge-enhanced version of the fluoroscopy frame. The algorithm includes a new multi-modal similarity measure and a novel technique for the calculation of the required gradients.
Three lower limb specimens were implanted with 1 mm tantalum beads to act as fiducial markers. Fluoroscopy data was captured for a knee flexion and femur and tibia CT data was registered to the fluoroscopy images.
A previous version of our algorithm (developed in 2008) showed good accuracy for in-plane translations and rotations of the knee bones. However, this algorithm did not have the ability to accurately determine out-of-plane translations. This lack of accuracy for out-of-plane translations has also been the major limitation of other single-plane 2D-3D registration algorithms. Fregly et. al. and Dennis et. al. reported standard deviations for this measurement of 5.6 and 3.03 mm respectively. The latest version of our algorithm achieves error standard deviations for out-of-plane translations of 0.65 mm. The algorithm includes a new similarity measure, which calculates the sum of the conditional variances (SCV) of the joint probability distributions of the images to be registered. This new similarity measure determines the true 3D position of the bones for a wider range of initial disparities and is also faster than the cross-cumulative residual entropy (CCRE) measure used in the 2008 version. For a set of initial 3D positions ranging from ± 5 pixels and ± 5 degrees the proposed approach successfully determined the correct 3D position for 96% of cases–whilst the approach using CCRE was successful for only 49% of cases. The algorithm also required 60% less iterations than the previous CCRE approach.
The new registration algorithm developed for the project provides a level of accuracy that is superior to other similar techniques. This new level of accuracy opens the way for a non-invasive mechanism for sophisticated kinematic analysis of knee joints. This will enable prospective, longitudinal and controlled studies of reconstruction surgery.
