Abstract
The limiting factor in the growth of RSA as a wide spread clinical tool is the man-hours needed to run a study. Calibration takes more than half of the processing time. The aim of this study is to develop automatic calibration method applicable to the grid and line patterns common in all RSA systems. This method uses a Harris Corner detector to find candidate positions on an image one 16th the original area (16 times quicker). Canny edge detection in regions of interest around the candidate positions on the full size image produce circular edges for marker-balls. A conic section is fitted to this edge using the Bookstein method to produce an accurate estimation of position to a local accuracy of 0.01 mm. Scanner distortion was modeled using a stabilised B-spline mesh to produce global accuracy of 0.03mm. A model based pattern recognition method can be used to label the marker-balls correctly. For sets of 4 marker balls a Homography was calculated and used to predict the positions of the other points in the grid. If supporting marker-balls are found in the predicted positions, they are counted. The four-point set, which returns the greatest number of support marker-balls, is the best estimate of a grid. Reference markers in the grid are used to localise it.
The method had a ninety- percent success rate on a set of 20 clinical X-rays. In two X-rays not enough marker-balls were visible due to a poor exposure. It finds marker-balls in a 15-MB image in 50 seconds on a 180 MHz silicon graphics O2. Labelling speed depends on the number of marker-balls and is 45 seconds per group of 50. This method is widely implementable, as it requires just the 3D positions of the markers in each plate of the calibration object for input.
The abstracts were prepared by Nico Verdonschot. Correspondence should be addressed to him at Orthopaedic Research Laboratory, University Medical Centre, PO Box 9101, 6500 HB Nijmegen, The Netherlands.
