I N:Springer Lecture Notes in Computer Science 1205 A Method For Evaluating CT-Based Surgical Registration R. E. Ellis D. J. Fleet J. T. Bryant J. Rudan P. Fenton Computing and Information Science Mechanical Engineering Surgery Radiology Queen’s University at Kingston, Ontario, Canada Abstract Workers in medical robotics and computer-assisted surgery are increasingly concerned with the accuracy of rigid registration of preoperative medical images to intraoperative data. Much of the work on evaluation methods is limited by a fundamental dilemma: either the method works only in vitro and provides absolute accuracy, or the method works in vivo but provides only relative accuracy. Previously, we developed a biocompatible fiducial marker and a CT-based detection technology that make it possible to evaluate the absolute in vivo accuracy of registra- tions. We have now developed a mathematical approach to evaluating accuracy, and have applied the method in a phantom study. 1 Introduction A system for computer-assisted orthopedic surgery, especially for total knee replace- ment, requires accurate registration of a preoperative 3D image to data collected intra- operatively. The accuracy required for knee replacement surgery has not been measured, but is generally accepted to be mm in position and in rotation. This estimate is bolstered by theoretical studies [2] and cadaver studies [5] that indicate misplace- ment of prosthetic components by only 2.5 mm severely affects the range of flexion and other kinematic variables. This paper presents an approach to evaluating registration accuracy, with results from an in vitro experimental application. Some current computer-assisted orthopedic systems [11] use large, invasive mark- ers that are implanted preoperatively under anæsthetic. A CT scan is then taken, and the images are processed semi-automatically to estimate the fiducial (marker) locations. During surgery the markers are typically located by a mechanical or optical pointing system, and registration is performed. Here, “registration” refers to the process of de- termining a rigid transformation from locations in pointer coordinates to locations of the same points in the volumetric image coodinate frame. Sources of registration errors include systematic errors within the pointing system and the imaging system, small random errors in the measuring systems, and errors within the registration algorithm. Calibration of the pointer system and imaging system should eliminate the systematic errors, but does not address the other error sources. Given measurements of fiducial locations in pointer coordinates and in image coor- dinates, one can compute a “nominal” registration. The ideal way to evaluate the regis- tration would be to use perfect knowledge (ground truth) of the locations of the fiducial points. From these true locations we could use models of measurement errors in the pointing and imaging systems to generate sets of alternative plausible measurements (e.g. by Monte Carlo simulation). Registrations can then be computed for many sets of plausible measurements. The resultant set of registrations then allows us to evaluate the stability of, and hence our confidence in, the computed nominal registration. Unfortunately, knowledge of the exact fiducial locations (perfect knowledge or ground truth) is not available. Instead, we use the original pointer and image measurements –