Quantitative validation of a new coregistration algorithm

A new coregistration software package, Neuro900 Image Coregistration software (Strichman Medical Equipment), has been developed specifically for nuclear medicine. With this algorithm, the correlation coefficient is maximized between volumes generated from sets of transaxial slices. No localization markers or segmented surfaces are needed. The coregistration program was evaluated for translational and rotational registration accuracy. A Tc-49m HM-PAO split-dose study (0.53 mCi low dose, L, and 1.01 mCi high dose, H) was simulated with a Hoffman Brain Phantom with five fiducial markers. Translation error was determined by a shift in image centroid, and rotation error was determined by a simplified two-axis approach. Changes in registration accuracy were measured with respect to: 1) slice spacing, using the four different combinations LL, LH, HL,, HH, 2) translational and rotational misalignment before coregistration, 3) changes in the step size of the iterative parameters. In all the cases the algorithm converged with only small difference in translation offset, /spl theta/ and /spl phi/. At 6 mm slice spacing, translational errors ranged from 0.9 to 2.8 mm (system resolution at 200 mm, 6.8 mm). The converged parameters showed little sensitivity to count density. In addition the correlation coefficient increased with decreasing iterative step size, as expected. From these experiments, we found that this algorithm based on the maximization of the correlation coefficient between studies was an accurate way to coregister SPECT brain images.<>