An automated image-registration technique based on multiple structure matching.

Abstract

A new image-registration technique that matches multiple structures on complementary imaging data sets (e.g., CT and MRI) has been developed and tested with both phantom and patient data. The algorithm assumes a rigid-body transformation and is suitable for correlating structures within the cranium or at the skull base. The basic premise of the new technique is that an optimum transformation is achieved when the relative volume lying outside of the intersection between a structure and its transformed counterpart is a minimum. This relative volume is calculated numerically using a random sampling approach, and a binary searching algorithm was used to step through the nine-dimensional parameter space consisting of three rotation angles, three scaling factors and three components of a translation vector. For the nine tests using phantom data, the automated structure-matching technique was able to predict the correct rotation angles to within +/- 1 degree. The expected clinical performance of the new technique was assessed by comparing results obtained with the new method to those obtained using other techniques for 12 patients who were treated with charged particles at Lawrence Berkeley Laboratory (LBL) and who had image-registration studies performed as part of their treatment plan. For 9 of the 12 patients considered, the new structure-matching technique produced a significantly better registration than the older methods, as measured by the resultant average relative volume lying outside of the intersection between any structure and its transformed counterpart. For the other three patients, results were not significantly different for the new structure-matching method and the older techniques.