MRI-SPET and SPET-SPET brain co-registration: evaluation of the performance of eight different algorithms

The aim of this study was to assess the accuracy and computing time needed for MRI-SPET and SPET-SPET brain co-registration using eight different algorithms (Hermes software from Nuclear Diagnostics Ltd run on a SUN Ultra Sparc 2) to determine the clinically most suitable algorithm. MRI-SPET co-registration was evaluated using phantom studies. To approximate clinical dual-headed SPET studies, a Hoffman brain phantom was filled with 99Tcm. For MRI imaging (1.5 Tesla), the phantom was filled with water and doped with Gd-DTPA for contrast enhancement. For both modalities, phantom images were acquired and reconstructed using a routine clinical protocol. MRI and SPET images were matched by Downhill Simplex minimization of the sum of absolute Count Differences (CD), the sum of the Square Root of absolute count differences (SR), the Difference in Shape between the binary masks (SD), the number of Sign Changes in the subtracted image (SC), the Variance of intensities between corresponding pixels (VAR), the sum of absolute count differences between the 2D- and 3D-Gradient images (2DG-3DG) and, finally, the standard deviation of the Uniformity Index (UI), that is the intensity ratio between spatially corresponding voxels. Six degrees of freedom were allowed (three translation and three rotation parameters, three scaling parameters were constrained). The accuracy of the matching process with these different similarity measures was evaluated via the residual mismatch between external markers. We found that CD, SR, VAR nad UI give the most accurate registration compared with the other similarity measures. For the evaluation of SPET-SPET co-registration, five 99Tcm-ECD brain perfusion SPET scans were performed with a dual-headed gamma camera. These studies were then manually misaligned, and subsequently re-aligned using the methods outlined above. For this application, CD, SR and VAR were also found to give the most accurate registration. For all of these algorithms, the computing time required was clinically acceptable (i.e. less than 10 min).