B-spline-based stereotactical normalization of brain FDG PET scans in suspected neurodegenerative disease: impact on voxel-based statistical single-subject analysis.

A b-spline-based method 'Lobster', originally designed as a general technique for non-linear image registration, was tailored for stereotactical normalization of brain FDG PET scans. Lobster was compared with the normalization methods of SPM2 and Neurostat with respect to the impact on the accuracy of voxel-based statistical analysis. (i) Computer simulation: Seven representative patterns of cortical hypometabolism served as artificial ground truth. They were inserted into 26 normal control scans with different simulated severity levels. After stereotactical normalization and voxel-based testing, statistical maps were compared voxel-by-voxel with the ground truth. This was done at different levels of statistical significance. There was a highly significant effect of the stereotactical normalization method on the area under the resulting ROC curve. Lobster showed the best average performance and was most stable with respect to variation of the severity level. (ii) Clinical evaluation: Statistical maps were obtained for the normal controls as well as patients with Alzheimer's disease (AD, n=44), Lewy-Body disease (LBD, 9), fronto-temporal dementia (FTD, 13), and cortico-basal dementia (CBD, 4). These maps were classified as normal, AD, LBD, FTD, or CBD by two experienced readers. The stereotactical normalization method had no significant effect on classification by of each of the experts, but it appeared to affect agreement between the experts. In conclusion, Lobster is appropriate for use in single-subject analysis of brain FDG PET scans in suspected dementia, both in early diagnosis (mild hypometabolism) and in differential diagnosis in advanced disease stages (moderate to severe hypometabolism). The computer simulation framework developed in the present study appears appropriate for quantitative evaluation of the impact of the different processing steps and their interaction on the performance of voxel-based single-subject analysis.