, Susan Resnick
, Christos DavatzikosJohns Hopkins School of Medicine Department of Radiology
Laboratory of Personality and Cognition, National Institute on Aging
The goal of this work is to test the accuracy and sensitivity of statistical parametric mapping (SPM) for detecting small or medium size brain activations from PET images, and to quantify the effect of two different spatial normalization methods in the detection of the activation foci. Computer simulated PET images were used.
High resolution MR images of 16 subjects in AC-PC orientation were segmented into gray matter (GM), white matter (WM), and CSF. Different radioactivity concentrations were uniformly assigned to GM, WM and CSF with ratios of 4:1:0.05 to mimic the radioactivity distribution during O-15 water studies of CBF. The effects of Poisson noise, scattering, and partial volume were modeled in the sinograms. The images were finally formed by filtered back projection using the scanner software. In order to model the limited axial field of view (FOV) of our scanner, we created 15 simulated slices 6.5 mm thick per subject. Six activation foci of spherical shapes with radii ranging from 5 to 10 mm were then planted into selected locations in the brain (temporal lobe, hippocampus, caudate, putamen, precentral knob, and cingulate gyri superior to the corpus callosum). We used our simulator to create baseline images, and images of two different activation levels with increased activity equal to 30% and 15%, respectively. These levels were assigned to the activation sites prior to the computer simulation of the formation of the PET images, and they correspond to roughly 15% and 7.5% in the final simulated PET images, due to noise and partial volume effects. Thus, we simulated a total of 16 baseline sets of images and 32 activated sets images (16 for each of the two activation levels).
We examined the effect of three independent factors on the detection of the activation sites: the activation level, the spatial normalization method, and the multiple comparison correction. We first used the SPM'96 spatial normalization method, and obtained unsatisfactory results, due to the limited (FOV) of our scanner. Therefore, the only two spatial normalization methods we compared were the one used by SPM'95, and the elastic registration described in [1]. The latter method used the high resolution MR images of these subjects to determine the spatial normalization transformation, and subsequently applied the same transformation to the PET images.
Our major conclusions are summarized next. For 30% activation (which corresponds to about 15% in the PET images), all but two (caudate and precentral gyrus) activation foci were accurately detected, with no false positives, when using the elastic registration method. The corrected p-value, however, for the caudate showed a trend toward significance (0.08). Only one focus was detected using the SPM spatial normalization (the cingulate gyrus, which was the largest site). For the 15% activation, only the two largest foci (cingulate gyri and temporal lobe) demonstrated significant activation when using the elastic registration. Only one (cingulate gyri) was detected using the SPM spatial normalization. No false positives were present when using the elastic normalization, while one false positive was observed when using SPM's normalization. When no multiple comparison correction was applied to the t-maps derived for a p=0.001 level threshold, and using a p=0.5 level extent threshold, all true activations were detected for both spatial normalization methods, and for both activation levels, although the hippocampus activation was not detected as a site separate from the right temporal lobe. This implies that the SPM voxel-wise correction of the p-values might be overly conservative. Overall, the most important conclusions drawn from these experiments are that more flexible spatial normalization methods can increase the sensitivity and accuracy of SPM quite remarkably, and that less conservative significance thresholds could better detect true activations, without necessarily increasing false positives. We note that our analysis applies only to the limited FOV data. Therefore, the SPM'96 spatial normalization method could possibly perform satisfactorily on datasets spanning the whole brain.
These simulation studies improve our understanding of the implications of statistical parametric maps, and assist in the interpretation of these maps. Moreover, they point to directions of future research for determining significance level thresholds and for developing new spatial normalization methods.
Reference
[1] C. Davatzikos, "Spatial Transformation and Registration of Brain Images Using Elastically Deformable Models", Special issue of CVIU on Med. Imag., 66(2):207-222, May 1997.
