NeuroImage 19 (2003) 601– 612
www.elsevier.com/locate/ynimg
Influence of the normalization template on the outcome of statistical parametric mapping of PET scans J.D. Gispert,a J. Pascau,a S. Reig,a R. Martı´nez-La´zaro,a V. Molina,b P. Garcı´a-Barreno,a and M. Descoa,* a
Medical Imaging Laboratory, Hospital Universitario “Gregorio Maran˜o´n,” Madrid, Spain b Department of Psychiatry, Hospital Doce de Octubre, Madrid, Spain Received 30 April 2002; revised 4 December 2002; accepted 23 December 2002
Abstract Spatial normalization is an essential preprocessing step in statistical parametric mapping (SPM)-based analysis of PET scans. The standard template provided with the SPM99 software package was originally constructed using 15O-H2O PET scans and is commonly applied regardless of the tracer actually used in the scans being analyzed. This work studies the effect of using three different normalization templates in the outcome of the statistical analysis of PET scans: (1) the standard SPM99 PET template; (2) an 18F-FDG PET template, constructed by averaging PET scans previously normalized to the standard template; and (3) an MRI-aided 18F-FDG PET template, constructed by averaging PET scans normalized according to the deformation parameters obtained from MRI scans. A strictly anatomical MRI normalization of each PET was used as a reference, under the rationale that a normalization based only upon MRI should provide higher spatial accuracy. The potential bias involved in the normalization process was estimated in a clinical SPM study comparing schizophrenic patients with control subjects. For each between-group comparison, three SPM maps were obtained, one for each template. To evaluate the influence of the template, these SPM maps were compared to the reference SPM map achieved using the anatomical normalization. SPMs obtained by MRI-aided normalization showed the highest spatial specificity, and also higher sensitivity when compared to the standard normalization using the SPM99 15O-H2O template. These results show that the use of the standard template under inappropriate conditions (different tracer or mental state) may lead to inconsistent interpretations of the statistical analysis. © 2003 Elsevier Science (USA). All rights reserved.
Introduction Statistical parametric mapping (SPM) is a method conceived to perform voxel-by-voxel statistical analysis of functional images (Friston et al., 1995). Spatial normalization is a required preprocessing step in intersubject statistical analysis that consists of applying the nonlinear deformations required to force every particular PET scan to match a reference template study. The algorithm minimizes the residual squared difference between the images being normalized and the template image (Ashburner and Friston, 1999). The main disadvantage of this approach lies in the
* Corresponding author. Unidad de Medicina y Cirugı´a Experimental, Hospital General Universitario “Gregorio Maran˜o´n,” Dr. Esquerdo, 46, E-28007 Madrid, Spain. Fax: ⫹34-426-51-08. E-mail address:
[email protected] (M. Desco).
total loss of natural or pathological variability in brain morphology. This problem might be of particular relevance when studying diseases like schizophrenia, known to involve changes in brain morphology (Lawrie and Abukmeil, 1998). Normalization can be performed directly by deforming the PET scans until they match the PET template or indirectly by using an additional MRI template. In this latter case, the deformation parameters are determined from structural images from the same subjects and then applied to the PET scans. This MRI-aided spatial normalization is allegedly more accurate than the one performed by using only functional images, given the better anatomical information and higher spatial resolution of MRI images (Ashburner and Friston, 1999). However, when MR images of the subjects under study are not available, it is only possible to perform a normal-
1053-8119/03/$ – see front matter © 2003 Elsevier Science (USA). All rights reserved. doi:10.1016/S1053-8119(03)00072-7
602
J.D. Gispert et al. / NeuroImage 19 (2003) 601– 612
ization based solely upon functional images. The SPM99 software package (Friston et al., 1999) includes a PET template that was originally created by averaging 12 15OH2O-PET scans from normal subjects with eyes closed in resting condition. For the normalization to work properly, the contrast in the template and the patient images must be reasonably similar. In PET scans, the image contrast is determined by both the tracer and the mental state of the subject under study. Therefore, when the acquisition conditions between the template and the PET scans differ, the dissimilarity of the image contrast may affect the outcome of the normalization process. Although these different acquisition conditions may not compromise the convergence of the normalization, their effect on the final result of the statistical analysis is uncertain. In the event of different acquisition conditions between the template and the actual PET scans, the construction of specific ad hoc templates has been proposed. One possible way of building a specific template consists of normalizing a set of 18F-FDG PET scans from a control group to the standard 15O-H2O PET template, and then averaging and smoothing these normalized images (Signorini et al., 1999). An alternative method to generate the template involves the use of anatomical MRI scans of the subjects in the control group. This method begins by co-registering the PET and MRI scans, then normalizing MRI scans to an MRI template, obtaining the deformation parameters that are finally applied to the control PET scans (Meyer et al., 1999). As in the previous method, these normalized scans are averaged and smoothed to generate the final template. In practice, however, most SPM analyses are performed by just using the standard template provided with the statistical software package SPM99, regardless of the tracer and the cerebral condition during the PET acquisition. The impact of different registration algorithms on the statistical analysis of neuroimages has been previously studied. For instance, Freire and Mangin (2001) proved that motion correction algorithms may introduce spurious activations in fMRI studies using both real and simulated time series with artificial activations. Regarding PET, great attention has been paid to the evaluation of the anatomical precision of different spatial normalization algorithms (Lancaster et al., 1999; Sugiura et al., 1999; Kochunov et al., 2000), whereas the possible effect of using different templates on the final statistical result has not been investigated in depth. Davatzikos et al. (2001) studied the effect of different normalization algorithms in the statistical outcome of SPM analysis by using PET phantom studies. They compared the methods available in SPM’95, SPM’96, and SPM99 with the STAR method (Davatzikos, 1997) that maps a parametric representation of the outer boundary of the brain and ventricles to the brain in the Talairach atlas (Talairach and Tournoux, 1988). Their results illustrate the importance of the normalization strategy, demonstrating that the use of anatomical MRI scans increases the sensitivity of the statistical results.
Ishii and colleagues (2001) examined the impact of two different brain normalization techniques: SPM99 and NEUROSTAT (Minoshima et al., 1994), on the metabolic patterns of Alzheimer’s disease patients as compared to healthy controls. Inconsistent results were found between these two normalization methods when applied to atrophied brains. In addition, they also showed that use of 18F-FDG and 15O-H2O PET templates for normalizing 18F-FDG PET scans resulted in different extent and peak height of areas representing metabolic changes. The aim of our study was to evaluate whether the choice of a particular normalization technique among those currently available in the literature could alter the clinical interpretation of SPM results in a realistic situation. Our setup is a case study comparing the statistical outcome of three different PET templates in an SPM analysis of schizophrenic patients and control subjects. Since we hypothesize that both the particular tracer used to construct the template and the potential anatomical alterations of the brain may bias the SPM analysis, three templates were chosen to encompass these factors: (1) the standard SPM99 PET template based on 15O-H2O; (2) an 18F-FDG PET template constructed by averaging FDG-PET scans previously normalized to the standard template (Signorini et al., 1999); and (3) an MRI-18F-FDG PET template constructed by averaging PET scans normalized using T1-weighted MRI scans (Meyer et al., 1999). A reference statistical map was calculated by using a normalization procedure that estimates the deformation parameters from MRI images of each subject. This procedure yields a pure anatomical normalization in which the different brain anatomy of the subjects is put into correspondence, in opposition to the functional normalization achieved by using PET templates, where only brain functional data are registered across subjects.
Materials and methods Subjects A test set of 18F-FDG PET and MR images was acquired from 17 normal subjects (CTRL) and 35 schizophrenic patients divided into two groups: 17 recent onset (RO) patients and 18 chronic (CHR) patients. Diagnosis was confirmed using the Structured Clinical Interview for DSM-IV (Diagnostic and Statistical Manual of Mental Disorders, fourth edition), using clinical interviews and information from families and clinical staff. Mean illness duration was 11.26 years (SD ⫽ 10.30) for the CHR patients and 1.82 years (SD ⫽ 1.92) in the RO group. CHR patients had shown a poor response to at least two different classical treatments, each lasting more than 1 month during the preceding year, with doses of at least 800 mg/day in CPZ equivalents. All the patients were on haloperidol at the moment of the PET scan. CHR patients received 10 –15 mg/day for 4 weeks before the PET scan and RO patients
J.D. Gispert et al. / NeuroImage 19 (2003) 601– 612
603
Fig. 1. (a) Standard SPM99 template; (b) FDG template built by averaging normalized 18F-FDG images to the standard 15O-H2O PET template; (c) MRI-FDG PET template built by using MRI-aided spatial normalization. Axial views are taken at the AC–PC plane and sagittal views at the interhemispheric plane.
received a minimal treatment of 5 mg/day during the 2 days preceding the study. After a complete description of the study to the subjects, written informed consent was obtained from each patient and from a first-degree relative. The research and ethical boards of the participant institutions endorsed the study. PET and MRI acquisition PET images were obtained on a Posicam HZL scanner (5.8 mm FWHM) 20 min after injection of 370 MBq of 18 F-FDG, in resting condition, eyes open in a dark room. Matrix size was 256 ⫻ 256 ⫻ 61 and slices were 2.6-mm thick. MRI volumes were obtained on a Philips ACS Gyroscan 1.5 T scanner, using a gradient echo T1-weighted 3D sequence (Flip angle ⫽ 30°, TR ⫽ 15.4 ms, TE ⫽ 4.6 ms) with matrix size 256 ⫻ 256 ⫻ 110 and voxel size 0.98 ⫻ 0.98 ⫻ 1.10 mm. Spatial normalization To assess the effect of using different templates in the normalization, two kinds of techniques were employed: functional normalization, which makes use of PET templates, and anatomical normalization, where the deformation is calculated solely upon MRI data. The outcome of SPM using three different procedures of calculating the PET template has been compared with that obtained after an MRI-based anatomical normalization, considered as the reference.
Functional normalization The three different functional templates were constructed as follows: (1) Standard SPM99 template (Fig. 1a): This is the standard 15O-H2O PET template provided with SPM99, and the most commonly used, even when other tracers are employed. This template was created by averging 12 PET scans from normal subjects with eyes closed in resting condition. Images were first registered to their corresponding T1-weighted MR images, and spatially transformed to the MNI (Montre´al Neurological Institute) reference space (Evans et al., 1993) using the transformation estimated from the anatomical scans. Original data for this template were acquired on a Siemens ECAT HR⫹, using Oxygen-15-labeled water. Averaged images were smoothed using an 8-mm FWHM Gaussian filter (Ashburner, 1999). (2) FDG template (Fig. 1b): Following the strategy described in Signorini et al. (1999), the PET scans of our 17 control subjects were normalized to the standard 15O-H2O PET template, using the algorithm provided with SPM99. The FDG PET template was built by averaging these normalized images and applying a smoothing Gaussian filter (FWHM ⫽ 8 ⫻ 8 ⫻ 8 mm). A remarkable difference with respect to the standard SPM template—apart from the different tracer—is that all control subjects were in resting condition with eyes opened during tracer uptake.
604
J.D. Gispert et al. / NeuroImage 19 (2003) 601– 612
Fig. 2. Result of the SPM analysis for the CHR vs CTRL comparison using the four different normalization schemes: (a) the standard SPM99 template; (b) the FDG template; (c) the MRI-FDG template; and (d) the reference anatomical normalization.
(3) MRI-FDG template (Fig. 1c): In this case, the PET template was built on the basis of previous MRIaided spatial normalization of the 17 control subjects, following the strategy described in Meyer et al. (1999). In order to build this template, PET scans of each control subject were registered to its corresponding MRI by using the MI (Mutual Information) maximization algorithm, included in the SPM99 package (Collignon et al., 1995). Co-registration can be performed accurately provided that only six rigidbody parameters are required (Ashburner and Friston, 1999). After this step, the MR images of the control subjects were normalized to the MRI template, also using the algorithm provided with SPM99, and the resulting deformation field was applied to the PET scans. As in the previous case, the final template was built by averaging these normalized PET images and applying a smoothing Gaussian filter (FWHM ⫽ 8 ⫻ 8 ⫻ 8 mm). This strategy is identical to the one followed to generate the standard SPM99 15O-H2O PET template. It is important to remark that MR images are only used in the stage of
template construction, not for the normalization of each patient scan. All the PET studies were normalized to these three functional templates with the algorithm provided by the SPM99 software package using its default options: 7⫻8⫻7 DCT basis functions, 12 nonlinear iterations, and the nonlinear regularization term set to medium. Bilinear interpolation was used in all the normalizations. All templates matched the standard MNI space (Evans et al., 1993). Anatomical normalization The normalization procedure used as reference followed the same strategy as the MRI-FDG template: the deformation parameters were obtained from the MRI normalization and then applied to the co-registered PET scan. However, in this case, each and every scan was normalized in this manner, without using any PET template. Consequently, the spatial normalization for the actual subjects is no longer performed by deforming the PET scans to match a PET template, as in the case of functional normalization. Instead, it is performed indirectly by co-registering the 17 control and the 35
J.D. Gispert et al. / NeuroImage 19 (2003) 601– 612
605
Table 1 Results of the CHR vs CTRL comparison using the standard SPM99 template Set level
Cluster level
Voxel level
P
c
Pcorr
Kc
Punc
Pcorr
T
Ze
Punc
0.003
6
0.000
4642
0.000
0.136 0.085
616 832
0.104 0.063
0.406 0.451 0.691
178 141 13
0.373 0.429 0.842
0.063 0.094 0.222 0.094 0.185 0.301 0.314 0.545 0.648
4.77 4.59 4.18 4.60 4.28 4.02 4.00 3.65 3.51
4.13 4.01 3.72 4.01 3.79 3.60 3.59 3.32 3.21
0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.001
x, y, z (mm)
Brain region
0, 48, ⴚ8 8, 42, 14 0, 20, ⫺12 ⴚ40, 14, 12 48, 12, 0 54, 20, 26 12, ⴚ64, 0 48, 48, 10 44, ⴚ60, 8
Ant. Cingulate
Right Insula Left Insula Primary Visual — —
Note. Voxel size, [2.0, 2.0, 2.0] mm (1 resel ⫽ 2547.78 voxels); search volumes, S ⫽ 1908648 mm3 ⫽ 238581 voxels ⫽ 87.4 resels; smoothness FWHM, 27.2 29.8 25.2 (mm) ⫽ 13.6 14.9 12.6 (voxels). Table shows local maxima ⬎ 8 mm apart per cluster, at voxel level uncorrected P ⬍ 0.001.
hypometabolic regions in the CHR schizophrenic patients and hypermetabolic regions in the RO patients. These particular statistical contrasts were chosen on the basis of previous evidence: the existence of a decreased frontal metabolism has been largely reported for classical neuroleptics such as haloperidol (Holcomb et al., 1996). Besides, another study (Kishimoto et al., 1998) found that thalamic and cingulate areas were hyperactive in 18F-FDG PET scans of acute schizophrenics similar to our RO patients in terms of illness and treatment conditions. Furthermore, associative frontal, parietal, and temporal gyri were hypoactive in a group of chronic schizophrenics, similar to our CHR patients. Following the most usual convention, SPM threshold was set to an uncorrected P value of P ⬍ 0.001 while significance threshold was set to a corrected P value of P ⬍ 0.05. These are the default and more usual thresholds when reporting results of SPM. According to this criterion a blob is accepted as “positive” whenever it includes at least one voxel with a corrected P value below 0.05. However, setting the proper thresholds to consider an observed P value as significant is critical in these types of studies (Andreasen,
schizophrenic MRI scans to the MRI template and then applying the estimated deformation field to the corresponding and previously co-registered PET scans. This normalization has been used as the reference, since the higher anatomical detail and spatial resolution of MRI with respect to PET imaging should theoretically provide more accurate spatial normalizations (Ashburner and Friston, 1999). SPM analysis The effect of the template on the SPM analysis was assessed by calculating the statistical parametric maps of two contrasts: CTRL vs CHR and CTRL vs RO. For each contrast, the three sets of results yielded by the three functional normalizations were compared against the results obtained using the anatomical normalization. PET images were analyzed with the SPM99 software package, using proportional scaling and a 20⫻20⫻20 mm FWHM Gaussian kernel smoothing. Gray-level threshold was set to 0.8; i.e., only voxels with an intensity level above 0.8 of the mean level for that scan were analyzed. Onetailed Student’s t tests were conducted to detect significant
Table 2 Results of the CHR vs CTRL comparison using the FDG template Set level
Cluster level
Voxel level
P
c
Pcorr
Kc
Punc
Pcorr
T
Ze
Punc
0.009
5
0.000
9318
0.000
0.033
1411
0.027
0.144 0.564 0.566
618 55 54
0.124 0.659 0.663
0.013 0.026 0.043 0.013 0.172 0.428 0.106 0.530 0.632 0.661
5.40 5.11 4.89 5.39 4.26 3.76 4.50 3.62 3.48 3.43
4.54 4.35 4.21 4.53 3.78 3.41 3.94 3.29 3.18 3.15
0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.001 0.001
x, y, z (mm)
Brain region
0, 46, ⴚ6 10, 44, 10 8, 10, ⫺20 52, 12, ⴚ4 38, 20, 6 56, 20, 24 ⴚ38, 14, 12 12, ⴚ60, 4 40, 56, 8 48, 50, 8
Ant. Cingulate
Left Insula
Right Insula Primary Visual —
Note. Voxel size, [2.0, 2.0, 2.0] mm (1 resel ⫽ 2874.89 voxels); search volume, S ⫽ 1924728 mm3 ⫽ 240591 voxels ⫽ 78.2 resels; smoothness FWHM, 28.3 30.7 26.4 (mm) ⫽ 14.2 15.4 13.2 (voxels). Table shows local maxima ⬎ 8 mm apart per cluster, at voxel level uncorrected P ⬍ 0.001.
606
J.D. Gispert et al. / NeuroImage 19 (2003) 601– 612
Table 3 Results of the CHR vs CTRL comparison using the MRI-FDG template Set level
Cluster level
Voxel level
P
c
Pcorr
Ke
Punc
Pcorr
T
Ze
Punc
0.108
3
0.000
10719
0.000
0.167 0.500
526 75
0.161 0.607
0.010 0.012 0.030 0.107 0.409
5.46 5.38 5.00 4.44 3.74
4.57 4.53 4.28 3.91 3.39
0.000 0.000 0.000 0.000 0.000
x, y, z (mm)
Brain region
ⴚ2, 48, ⴚ4 46, 14, 2 4, 12, ⫺14 ⴚ38, 16, 16 10, ⴚ58, 10
Ant. Cingulate Left Insula Right Insula Primary Visual
Note. Voxel size, [2.0, 2.0, 2.0] mm (1 resel ⫽ 2874.89 voxels); search volume, S ⫽ 1797056 mm3 ⫽ 224632 voxels ⫽ 70.5 resels; smoothness FWHM, 28.9 31.6 26.2 (mm) ⫽ 14.4 15.8 13.1 (voxels). Table shows local maxima ⬎ 8 mm apart per cluster, at voxel level uncorrected P ⬍ 0.001.
Direct subtractions of these SPM maps are also provided. Red and blue regions indicate higher and lower t values, respectively, for each particular map with respect to the reference.
1996). For this reason, we carried out a validation procedure to determine the adequate level of significance for our data. Assuming that no significant differences should be expected within a control group, we used a bootstrap resampling technique to extract 2000 random subgroups on which we performed the same tests as on the patient groups (Efron and Tibshirani, 1986). The four normalization methods were individually tested using this technique, showing little differences between runs and thus providing an empirical validation of the robustness of the study. False positive ratios of the SPM maps at the level of corrected P value of P ⬍ 0.05 were: 0.048 with the standard SPM99 template, 0.046 with the FDG template, 0.036 with the MRI-FDG template, and 0.041 with the anatomical normalization. To better depict differences among SPM results, each of the three SPM maps were subtracted from the reference SPM result, revealing the brain regions where the use of different normalization templates results in noticeable disagreements.
Chronic patients (CHR) vs control subjects (CTRL) The MIPs for the four SPM maps corresponding to this comparison are presented in Fig. 2. P values of the hypometabolic brain regions of CHR patients are detailed in Tables 1– 4. Two regions (anterior cingulate and left insula) appeared as significant when using the reference anatomical normalization. When using functional normalization methods, all suprathreshold blobs show larger spatial extent and, despite the fact that the statistical maps appear very similar, the standard SPM99 template failed to detect as significant the anterior cingulate and the left insula at the level of corrected P ⬍ 0.05. The primary visual area appears in all the statistical maps corresponding to functional normalization methods (uncorrected P ⬍ 0.001) but not in the map calculated with the reference normalization. Fig. 3 shows the difference images between the SPM maps obtained using the three PET templates and the MRI normalization for the CHR vs CTRL comparison. The Parieto-Occipital region showed lower P values in all the statistical maps calculated by using functional normalization as compared to the reference anatomical normalization.
Results SPM results comparing the two different patient groups vs the same control group are presented. Each comparison between controls and patients yielded four different SPMs, three corresponding to the functional templates plus the MRI normalization. Results are shown on maximum intensity projections (MIP). Table 4 Results of the CHR vs CTRL comparison using the anatomical normalization Set level
Cluster level
Voxel level
P
c
Pcorr
Ke
Punc
Pcorr
T
Ze
Punc
0.065
4
0.001
3167
0.000
0.022
1234
0.015
0.199
379
0.149
0.695
16
0.794
0.012 0.030 0.360 0.035 0.079 0.258 0.171 0.564 0.720
5.49 5.12 3.95 5.05 4.70 4.14 4.34 3.66 3.44
4.60 4.36 3.55 4.32 4.09 3.69 3.84 3.33 3.16
0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.001
x, y, z (mm)
Brain region
ⴚ2, 52, ⴚ2 ⫺4, 36, 26 2, 16, ⫺10 44, 22, 28 44, 16, 2 42, 34, 18 ⴚ44, 16, 4 ⫺46, 14, 26 62, ⴚ18, 4
Ant. Cingulate
Left Insula
Right Insula —
Note. Voxel size, [2.0, 2.0, 2.0] mm (1 resel ⫽ 2023.40 voxels); search volume: S ⫽ 1643296 mm3 ⫽ 205412 voxels ⫽ 94.5 resels; smoothness FWHM, 24.4 28.3 23.4 (mm) ⫽ 12.2 14.1 11.7 (voxels). Table shows local maxima ⬎ 8 mm apart per cluster, at voxel level uncorrected P ⬍ 0.001.
J.D. Gispert et al. / NeuroImage 19 (2003) 601– 612
607
Fig. 3. Difference between statistical maps obtained in the CHR vs CTRL comparison with respect to the reference anatomical normalization: (a) the standard SPM99 template; (b) the FDG template; (c) the MRI-FDG template. Red/blue regions indicate differences higher/lower than 2 in the t value (33 df) for each particular map with respect to the reference.
However, higher P values were found in the primary visual area when using standard SPM99 normalization, and in the anterior cingulate and the pons across all statistical maps. Recent onset (RO) vs CTRL Fig. 4 shows the MIP for the RO vs CTRL comparison: Fig. 4a, b, and c correspond to the normalization performed with the standard SPM99, FDG, and MRI-FDG templates and Fig. 4d to the reference procedure. P values of the hypermetabolic regions of RO patients are shown in Tables 5– 8. The RO vs CTRL comparison when using anatomical normalization (Fig. 4d) detected no differences except for a nonsignificant spot in the parietal region. None of the regions appearing in any SPM obtained with functional nor-
malizations reached the significance threshold. However, their spatial extent becomes progressively larger when using the MRI-FDG, FDG, and standard templates, as compared to the reference. Fig. 5 shows the difference images between the SPM maps obtained using the three PET templates and the MRI normalization for the RO vs CTRL comparison. The direct subtraction corresponding to the standard SPM99 method showed the higher mismatch as compared to the reference anatomical normalization. Higher P values were located in the parietal, temporal, and frontal lobes, whereas both left and right insulas and the basal ganglia showed lower P values. With regard to the maps obtained when using the two FDG templates, no significant difference was found with respect to the reference method.
608
J.D. Gispert et al. / NeuroImage 19 (2003) 601– 612
Fig. 4. Result of the SPM analysis for the RO vs CTRL comparison using the four different normalization schemes: (a) the standard SPM99 template; (b) the FDG template; (c) the MRI-FDG template; and (d) the reference anatomical normalization.
Discussion In this work we studied the effect of several normalization procedures frequently employed by the SPM users community, in a typical setting of patients vs controls comparison intended to be representative of a variety of studies in neuroscience and psychiatric research. All the normalization and co-registration operations have been performed with the algorithms provided with the software package SPM99 for this purpose, thus facilitating the replication of our study. The discussion focuses on the differences between the SPM results obtained with the different normalization schemes. The FDG and MRI-FDG templates are discussed together, since they achieved very similar results (Figs. 2b and c, 4b and c, and Tables 2, 3, 6, and 7). Standard SPM99 template vs anatomical normalization In the CHR vs CTRL comparison, the statistical map obtained with the standard SPM99 template (Fig. 2a and
Table 1) did not detect statistically significant regions (corrected P value ⬍ 0.05). However, two regions (anterior cingulate and left insula) reached the significance threshold when using the reference anatomical normalization for the same comparison (Fig. 2d and Table 4). This suggests that the use of the standard SPM99 template yields less sensitive results when compared to those obtained by the anatomical normalization. In this regard, our results obtained in a practical setting confirm experimental work using phantom PET scans (Davatzikos et al., 2001), that also reported a lower sensitivity of the standard SPM normalization. Primary visual area did not reach the significance threshold either in the CHR vs CTRL or in the RO vs CTRL comparisons. However, this region showed larger spatial extent when using the standard template than when using the reference normalization (Figs. 2d and 4d). Accordingly, the direct subtraction of the probability maps showed discrepancy in the occipital lobe (Fig. 3a). This could be explained by the contrast mismatch between the scans and
J.D. Gispert et al. / NeuroImage 19 (2003) 601– 612
609
Table 5 Results of the RO vs CTRL comparison using the standard SPM99 template Set level
Cluster level
Voxel level
x, y, z (mm)
P
c
Pcorr
Ke
Punc
Pcorr
T
Ze
Punc
0.000
11
0.466 0.558
131 73
0.444 0.577
0.603 0.677 0.655 0.551 0.699 0.729 0.734 0.740 0.746
50 19 27 77 12 4 3 2 1
0.652 0.799 0.753 0.566 0.848 0.924 0.937 0.951 0.969
0.106 0.376 0.395 0.446 0.472 0.558 0.614 0.615 0.689 0.690 0.693 0.719 0.743
4.57 3.91 3.88 3.81 3.77 3.65 3.57 3.57 3.47 3.46 3.46 3.42 3.39
3.98 3.51 3.49 3.43 3.40 3.31 3.25 3.25 3.17 3.17 3.16 3.14 3.11
0.000 0.000 0.000 0.000 0.000 0.000 0.001 0.001 0.001 0.001 0.001 0.001 0.001
46, 44, 48, 32, 16, ⴚ14, 44, 38, ⴚ22, 12, 44, ⴚ48, ⴚ50,
ⴚ62, 46, 50, 54, 24, 38, ⴚ8, ⴚ38, 10, ⴚ78, 32, 16, 14,
Brain Region
60 ⴚ22 ⫺16 ⫺22 68 58 34 48 72 60 ⴚ28 ⴚ26 ⴚ24
Parieto-Occipital Left Orbital
— Right Prefrontal Parietal — — — — Right Sup. Temp. —
Note. Voxel size, [2.0, 2.0, 2.0] mm (1 resel ⫽ 2531.26 voxels); search volume: S ⫽ 1925160 mm3 ⫽ 240645 voxels ⫽ 89.0 resels; smoothness FWHM, 26.9 30.1 25.1 (mm) ⫽ 13.4 15.0 12.5 (voxels). Table shows local maxima ⬎ 8 mm apart per cluster, at voxel level uncorrected P ⬍ 0.001.
the template, as a result of their different radiotracer and mental conditions (open/closed eyes). The larger spatial extent of the blobs obtained when using the standard SPM99 template suggests a lower spatial specificity of this method, as compared to the anatomical normalization. This could be explained by the higher resolution of the anatomical images which may allow for a normalization at a finer scale. The direct subtraction of probability maps, corresponding to the standard SPM99 method in the RO vs CTRL comparison, presented the higher mismatch among all cases (Fig. 5). This contrast mismatch can
be interpreted as an indicator of lack of registration accuracy, thus pointing to the standard SPM99 method as the most inaccurate registration in this case. However, these mismatches did not alter the clinical interpretation of the SPM result, since none of the areas involved reached the P threshold. This may support the lack of a direct relation between the normalization accuracy achieved and its effect on the final SPM results. Our findings are consistent with those of Ishi and colleagues (2001), who also found that differently normalized images can still yield very similar SPMs.
Table 6 Results of the RO vs CTRL comparison using the FDG template Set level
Cluster level
Voxel level
x, y, z (mm)
P
c
Pcorr
Ke
Punc
Pcorr
T
Ze
Punc
0.000
17
0.426 0.438 0.551
146 135 55
0.464 0.482 0.670
0.672 0.664 0.647 0.654 0.676 0.639 0.687 0.681 0.681 0.687 0.687 0.687 0.687 0.687
4 6 11 9 3 14 1 2 2 1 1 1 1 1
0.933 0.913 0.872 0.887 0.944 0.851 0.973 0.957 0.957 0.973 0.973 0.973 0.973 0.973
0.178 0.281 0.373 0.461 0.656 0.444 0.511 0.523 0.573 0.605 0.636 0.658 0.662 0.669 0.670 0.685 0.688 0.692 0.695
4.24 4.00 3.84 3.70 3.42 3.73 3.63 3.61 3.54 3.50 3.45 3.42 3.42 3.41 3.41 3.38 3.38 3.37 3.37
3.75 3.58 3.45 3.35 3.14 3.37 3.30 3.28 3.23 3.19 3.16 3.14 3.13 3.12 3.12 3.10 3.10 3.10 3.09
0.000 0.000 0.000 0.000 0.001 0.000 0.000 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001
ⴚ46, 42, 46, 36, 40, ⴚ36, 44, ⴚ20, 22, 46, 54, 42, 44, 8, 50, ⴚ14, 44, ⴚ16, 44,
20, ⴚ6, 46, 48, 40, 52, ⴚ64, 40, ⴚ48, ⴚ68, ⴚ52, ⴚ88, ⴚ60, ⴚ76, 48, 38, ⴚ88, 34, ⴚ86,
Brain region
⫺28 32 ⴚ20 ⫺24 ⫺26 30 60 54 84 56 28 28 62 62 ⴚ16 58 24 60 30
Right Sup. Temp. Parietal Left Orbital
— Parieto—Occipital Right Prefrontal — — — — — — — — — — —
Note. Voxel size, [2.0, 2.0, 2.0] mm (1 resel ⫽ 3067.31 voxels); search volume, S ⫽ 1930536 mm3 ⫽ 241317 voxels ⫽ 73.7 resels; smoothness FWHM, 28.8 31.8 26.8 (mm) ⫽ 14.4 15.9 13.4 (voxels). Table shows local maxima ⬎ 8 mm apart per cluster, at voxel level uncorrected P ⬍ 0.001.
610
J.D. Gispert et al. / NeuroImage 19 (2003) 601– 612
Table 7 Results of the RO vs CTRL comparison using the MRI-FDG template Set level
Cluster level
Voxel level
P
c
Pcorr
Ke
Punc
Pcorr
T
Ze
Punc
0.000
7
0.613 0.623 0.599 0.623 0.643 0.649 0.649
10 7 15 7 2 1 1
0.883 0.906 0.849 0.906 0.958 0.973 0.973
0.471 0.494 0.557 0.625 0.641 0.652 0.655
3.63 3.60 3.51 3.41 3.39 3.38 3.37
3.30 3.27 3.20 3.13 3.11 3.10 3.09
0.000 0.001 0.001 0.001 0.001 0.001 0.001
x, y, z (mm)
Brain region
42, ⴚ58, 64 ⴚ22, 44, 54 38, ⴚ2, 36 ⴚ48, 16, ⴚ20 ⴚ24, 38, 58 38, ⴚ54, 68 ⴚ16, 44, 56
Parieto-Occipital Prefrontal (R) Parietal Orbital (L) — — —
Note. Voxel size, [2.0, 2.0, 2.0] mm (1 resel ⫽ 3203.27 voxels); search volume, S ⫽ 1791528 mm3 ⫽ 223941 voxels ⫽ 65.5 resels; smoothness FWHM, 29.3 32.7 26.8 (mm) ⫽ 14.7 16.3 13.4 (voxels). Table shows local maxima ⬎ 8 mm apart per cluster, at voxel level uncorrected P ⬍ 0.001.
FDG & MRI-FDG templates vs anatomical normalization In the CHR vs CTRL comparison, both the anterior cingulate and the left insula appear as significant (Figs. 2b and c), in agreement with the results obtained using anatomical normalization (Fig. 2d). This implies that neither of the FDG templates caused, in our dataset, the loss of sensitivity found with the standard SPM99 template. None of the regions appearing in the RO vs CTRL comparison reached the significance threshold (Figs. 4b and c), thus corroborating the higher spatial specificity of the anatomical normalization with respect to the rest of the methods. The fully anatomical normalization ensures a total independence of the registration and statistical analysis procedures, since each one is performed using different data: MRI for the registration and PET for the analysis. This is not the case when using only functional images, since voxel information from the PET scans is used for both the spatial normalization and the statistical analysis. Therefore, an adverse interaction may exist between the data used for the construction of templates and the data being actually analyzed, as previously suggested by Bookstein (2001). This interaction could explain the loss of spatial accuracy and/or statistical power when SPM studies are solely based on functional templates. Intersubject registration of 18F-FDG PET scans to the standard 15O-H2O SPM99 template seems to introduce bias in SPM studies, probably because of the contrast disparity. Use of templates generated from images obtained in the same mental condition and based on the same tracer, like the FDG and MRI-FDG templates described in this paper, may provide an alternative solution to avoid that bias.
A limitation of our work is that the conclusions derive from a case study, though, on the other hand, our assessment has interesting practical implications as the setting is quite representative of those commonly employed in psychiatry research. Nevertheless, a thorough evaluation of intersubject registration methods with real patient data is very hard due to the lack of a ground-truth to compare with. Our study does not intend to demonstrate that a fully MRI-based (or anatomical) normalization is the definitive method, since it has been selected a priori as a reference, on the basis of previous literature (Ashburner and Friston, 1999; Davatzikos et al., 2001).
Conclusions Our results indicate that the use of different normalization strategies may alter noticeably the SPM maps, even leading to a different clinical interpretation. Normalizing procedures using templates that differ from the PET scans in tracer or mental condition achieved a lower sensitivity when compared to specific templates without this source of error. When constructing these templates, the use of additional MRI anatomical information did not seem to improve their accuracy substantially. All functional normalization methods failed to reach the high spatial specificity achieved by the reference anatomical registration. Specific templates can be easily constructed from control scans, using the algorithms provided by the standard SPM99 software package. Normalization using these templates showed a statistical outcome similar to that obtained using the fully anatomical normalization, which requires
Table 8 Results of the RO vs CTRL comparison using the anatomical normalization Set level P —
Cluster level c —
Voxel level
Pcorr
Ke
Punc
Pcorr
T
Ze
Punc
0.511
5
0.890
0.515
3.42
3.13
0.001
x, y, z (mm)
Brain region
36, 0, 36
Parietal
Note. Voxel size, [2.0, 2.0, 2.0] mm (1 resel ⫽ 1772.81 voxels); search volume, S ⫽ 702160 mm3 ⫽ 87770 voxels ⫽ 45.0 resels; smoothness FWHM, 23.5 25.8 23.4 (mm) ⫽ 11.7 12.9 11.7 (voxels). Table shows local maxima ⬎ 8 mm apart per cluster, at voxel level uncorrected P ⬍ 0.001.
J.D. Gispert et al. / NeuroImage 19 (2003) 601– 612
611
Fig. 5. Difference between statistical maps obtained in the RO vs CTRL comparison with respect to the reference anatomical normalization: (a) the standard SPM99 template; (b) the FDG template; and (c) the MRI-FDG template. Red/blue regions indicate differences higher/lower than 2 in the t value (32 df) for each particular map with respect to the reference.
MRI data from all the subjects in the study. The widespread practice of using templates based on tracers other than the one actually used in the PET scans should be strongly discouraged.
Acknowledgments This work was supported by Grants FIS-00/0036, Comunidad de Madrid-III PRICIT, and Fundacio´ La Caixa (99/ 042-00). The authors thank Dr. Miguel Angel Pozo from Centro PET Complutense de Madrid for his collaboration in the acquisition of the PET scans and Dr. Celso Arango from the Psychiatry Department of Hospital General Universitario “Gregorio Maran˜o´n” for his valuable comments.
References Andreasen, N.C., 1996. PET and the [15O]H2O technique, Part 2: Choosing a significance threshold. Am. J. Psychiatr. 153, 6. Ashburner, J., 1999. SPM99 Manual. Ashburner, J., Friston, K.J., 1999. Nonlinear spatial normalization using basis functions. Hum. Brain Mapping 7, 254 –266. Bookstein, F.L., 2001. “Voxel-based morphometry” should not be used with imperfectly registered images. NeuroImage 14, 1454 – 1462. Collignon, A., Maes, F., Delaere, D., Vandermeulen, D., Suetens, P., Marchal, G., 1995. Automated multi-modality image registration based on information theory, in: Information Processing in Medical Imaging, Kluwer Academic Publishers, Amsterdam. Davatzikos, C., 1997. Spatial transformation and registration of brain images using elastically deformable models. Comput. Vis. Image Underst. 66, 207–222.
612
J.D. Gispert et al. / NeuroImage 19 (2003) 601– 612
Davatzikos, C., Li, H.H., Herskovits, E., Resnick, S.M., 2001. Accuracy and sensitivity of detection of activation foci in the brain via statistical parametric mapping: a study using a PET simulator. NeuroImage 13, 176 –184. Efron, B., Tibshirani, R., 1986. Bootstrap methods for standard errors, confidence intervals, and other measures of statistical accuracy. Stat. Sci. 1, 54 –77. Evans, A.C., Collins, D.L., Mills, S.R., Brown, E.D., Kelly, R.L., Peters, T.M., 1993. 3D statistical neuroanatomical models from 305 MRI volumes, in: IEEE-Nuclear Science Symposium and Medical Imaging Conference, Piscataway, NJ, pp. 1813–1817. Freire, L., Mangin, J.F., 2001. Motion correction algorithms may create spurious brain activations in the absence of subject motion. NeuroImage 14, 709 –722. Friston, K., Ashburner, J., Heather, J., Holmes, A., Poline, J.B., 1999. Statistical Parametric Mapping. The Wellcome Department of Cognitive Neurology. University College London, London. (http://www. fil.ion.ucl.ac.uk/spm). Friston, K.J., Holmes, A.P., Worsley, K.J., Poline, J.P., Frith, C.D., Frackowiak, R.S.J., 1995. Statistical parametric maps in functional imaging: a general linear approach. Hum. Brain Mapping 2, 189 –210. Holcomb, H.H., Cascella, N.G., Thaker, G.K., Medoff, D.R., Dannals, R.F., Tamminga, C.A., 1996. Functional sites of neuroleptic drug action in the human brain: PET/FDG studies with and without haloperidol. Am. J. Psychiatr. 153, 41– 49. Ishii, K., Willoch, F., Minoshima, S., Drzezga, A., Ficaro, E.P., Cross, D.J., Kuhl, D.E., Schwaiger, M., 2001. Statistical brain mapping of 18FFDG PET in Alzheimer’s disease: validation of anatomic standardization for atrophied brains. J. Nucl. Med. 42, 548 –557.
Kishimoto, H., Yamada, K., Iseki, E., Kosaka, K., Okoshi, T., 1998. Brain imaging of affective disorders and schizophrenia. Psychiatr. Clin. Neurosci. 52 (Suppl), S212–S214. Kochunov, P., Lancaster, J., Thompson, P., Boyer, A., Hardies, J., Fox, P., 2000. Evaluation of octree regional spatial normalization method for regional anatomical matching. Hum. Brain Mapping 11, 193–206. Lancaster, J.L., Fox, P.T., Downs, H., Nickerson, D.S., Hander, T.A., El Mallah, M., Kochunov, P.V., Zamarripa, F., 1999. Global spatial normalization of human brain using convex hulls. J. Nucl. Med. 40, 942–955. Lawrie, S.M., Abukmeil, S.S., 1998. Brain abnormality in schizophrenia. A systematic and quantitative review of volumetric magnetic resonance imaging studies. Br. J. Psychiatr. 172, 110 –120. Meyer, J.H., Gunn, R.N., Myers, R., Grasby, P.M., 1999. Assessment of spatial normalization of PET ligand images using ligand-specific templates. NeuroImage 9, 545–553. Minoshima, S., Koeppe, R.A., Frey, K.A., Kuhl, D.E., 1994. Anatomic standardization: linear scaling and nonlinear warping of functional brain images. J. Nucl. Med. 35, 1528 –1537. Signorini, M., Paulesu, E., Friston, K., Perani, D., Colleluori, A., Lucignani, G., Grassi, F., Bettinardi, V., Frackowiak, R.S., Fazio, F., 1999. Rapid assessment of regional cerebral metabolic abnormalities in single subjects with quantitative and nonquantitative [18F]FDG PET: a clinical validation of statistical parametric mapping. NeuroImage 9, 63– 80. Sugiura, M., Kawashima, R., Sadato, N., Senda, M., Kanno, I., Oda, K., Sato, K., Yonekura, Y., Fukuda, H., 1999. Anatomic validation of spatial normalization methods for PET. J. Nucl. Med. 40, 317–322. Talairach, J., Tournoux, P., 1988. A Co-planar Stereotactic Atlas of the Human Brain. Thieme, New York.
