Dose Dependent Effects of the CRF1 Receptor Antagonist R317573 on Regional Brain Activity in Healthy Male Subjects

Schmidt, MES et al.

Information on the PET image processing procedures

Randolph D. Andrews1, Terry Brown1, Koen Van Laere2

1Abiant, Inc, Deerfield, IL, USA

2Division of Nuclear Medicine, University Hospital and K.U.Leuven, Leuven, Belgium

Correspondence on PET image processing procedures:

Randolph D. Andrews

Abiant, Inc.

420 Lake Cook Road

Suite 111

Deerfield, Illinois 60015

USA

Phone: +1 847 282 3813

Email:

Spatial normalization

Statistical Parametric Mapping-2 (SPM2) was used to perform spatial normalization. This normalization was first done by calculating an affine registration with 12 degrees of freedom—rotation, translation, and shear—to match the individual anatomy to the anatomy of a given template, then computing a non-linear spatial warping using low-frequency basis functions to provide an even closer match to the template, mediated by a cost function.[1] A Positron Emission Tomography (PET) template specific to this study was created by normalizing the PET scans with the aid of individual subjects’ Magnetic Resonance Images (MRIs) and using the MRI-aligned PET scans to create a template.[2] The MRI co-registration was performed by first realigning each individual PET scan with the other scans from the same subject. Spatial normalizations were calculated to match the individuals’ MRIs to the Montreal Neurological Institute (MNI) T1 MRI template. The calculated spatial transforms were applied to each coregistered PET scan to bring it into standard MNI space. These transformed PET scans were averaged to create a grand mean, producing a new template that was unique to both the condition and radiotracer of this study. The realigned PET scans were again normalized to this unique template. The PET images were resliced to 2 mm voxels.

Smoothing

In order to prepare the data for subsequent inferences derived from the Gaussian Random Field Model and enhance the signal to noise ratio of the image set, a Gaussian filter was applied. In FDG-PET studies of less than 40 subjects examining drug action it is critical to minimize the effect of individual scan outliers that may arise due to preprocessing variables such as co-registration, spatial normalization and smoothing in subsequent inferential tests.[3] The choice of a single threshold at the midpoint between 2 group means is optimal only if both population distributions are Gaussian with the same full-width at half-maximum (FWHM).[4] In practice, this assumption is rarely satisfied if the selection of a lesser smoothing kernel, while achieving a larger or equal overall signal-to-noise rations (SNR), also retains individual subjects whose first principal component is not normally distributed.

As part of selecting the optimal set of pre-processing parameters for this data set, two separate principal components analyses (PCAs) were used to examine SNRs in the dataset and to inspect the contribution of individual scans to the major components. A population-based (blinded to treatment) PCA was used to identify scans that differ more than three standard deviations from the population mean, while a state-based PCA was used to identify individual scan characteristics in response to a specific drug condition that are more than three standard deviations from the population mean if that subject is removed. The population PCA may be seen to identify gross outliers due to poor spatial normalization, and as such, regularization and cutoff parameters may be adjusted until the assumption of a general Gaussian distribution is satisfied.[5] In this study, the SPM2 default parameters for spatial normalization (medium regularization and 25mm cutoff) satisfied the inclusion criteria above, so the effect of increasing smoothing kernel size was evaluated to identify outliers due to inter-individual differences in drug response. Figure 1 shows the starting profile of the general population of scans (blinded) and the effect of smoothing on the empirical SNR of the first six principal components.

Figure 1 Effect of Smoothing on Individual PCs Empirical SNR (Population-Based PCA, Blinded to Treatment)


With regard to the state-based PCA, however, at smaller smoothing kernels (8mm and 12mm), one subject retained outlier status until images were smoothed with a 16mm filter (see Table 1 in Attachment 1). As a result, a 16 mm kernel size was used for this data set. It was interesting to note that while the empirical SNR for the first principal component in the population-based PCA decreased as expected (eg., from 4.570 to 4.121, 12mm and 16mm smoothing kernels respectively) the SNR for PC1 actually increased slightly (from 4.186 to 4.189) with the implementation of the higher smoothing kernel in the state-based PCA (i.e., after segregation of scans into their respective treatment conditions).

Masking

A voxel exclusion mask was generated and applied to the spatially normalized images to prevent biasing image data due to ventricles, white matter, and regions outside the brain. A high pass filter based on the grand mean image (the mean of all scans irrespective of treatment, prior to smoothing) was used to remove the skull, white matter,

and ventricular space on all images. An intensity threshold of 40% was applied to the grand mean image (thus retaining the upper 60% of voxels by intensity) and subsequently binarized. No slices in the spatially normalized images fell outside the field of view, so all voxels (upper 60%) were retained for this analysis.

Intensity (Z-score) Normalization

To further minimize physiological biases in the analysis, the masked image data was normalized relative to whole brain gray matter mean and standard deviation z-scores for each voxel. [6] The resultant images with mean value of zero were entered as data into SPM analysis.

Data analysis

In order to identify the areas with significant change in activity in relation to the experimental parameters, SPM was used to assess the significance of the changes, computed voxel by voxel, resulting in a statistical parametric map. Image data were analyzed using SPM2 (The Wellcome Department of Cognitive Neurology, London, UK; Statistical analyses were performed with the multi-subject: conditions & covariates option (3 conditions, +0 covariate, +8 block, +0 nuisance) of the PET models design class and applying the non-sphericity estimation option which enables multivariate analyses within a univariate framework. Statistical tests involved locating regionally specific differences in gray matter between treatment groups in accordance with the SPM general linear model. [7] Clusters were initially identified by height thresholds given at significance levels of p < 0.01 (uncorrected) with an extent threshold of 50 voxels. Individual (orthogonal) SPM contrasts were then corrected for multiple comparisons by employing false discovery rate (FDR) correction with the FDR rejection level set at p<0.05. Anatomical labels were obtained from the coordinates given in the SPM results via submission to the Talairach Daemon ( The coordinates of these cluster maxima were converted from MNI to Talairach space using the standard linear conversion algorithm, mni2tal. [8]

To identify conjoint areas of significant activation in differing drug doses, we employed a Boolean AND routine that compares the paired t-tests (treatment-placebo) from two different treatments using the placebo as the common parameter. This procedure identifies clusters where both treatment images intersect with placebo at specified height and extent thresholds derived from the execution of FDR-correction.

References

  1. Ashburner J, Friston KJ (1999) Nonlinear spatial normalization using basis functions. Human Brain Mapping 7(4):254-266.
  2. Gispert JD, Pascau J, Reig S, Martinez R, Molina V, Desco M. (2002): Effect of the normalization template in statistical parametric mapping of PET scans. Proc IEEE International Symposium on Biomedical Imaging. 851-854.
  3. Van Horn JD, Ellmore TM, Esposito G, Berman KF (1998): Mapping voxel-based statistical power on parametric images. Neuroimage 7(2):97-107.
  4. Liow JS, Rehm K, Strother SC, Anderson JR, Morch N, Hansen LK, Schaper KA, Rottenberg DA. (2000): Comparison of voxel- and volume-of-interest-based analyses in FDG PET scans of HIV positive and healthy individuals. J Nucl Med. 41(4):612-21.
  5. Lukic AS, Wernick MN, Yang Y, Hansen LK, Arfanakis K, Strother SC. (2007): Effect of spatial alignment transformations in PCA and ICA of functional neuroimages. IEEE Trans Med Imaging 26(8):1058-68.
  6. McIntosh AR, Grady CL, Haxby JV, Maisog J.Ma, Horwitz B, Clark CM (1996): Within-subject transformations of PET regional cerebral blood flow data: ANCOVA, ratio, and Z-score adjustments on empirical data. Human Brain Mapping 4(2): 93-102.
  7. Friston KJ, Frith CD, Liddle PF, Frackowiak RS. (1991): Comparing functional (PET) images: the assessment of significant change. J Cereb Blood Flow Metab. Jul;11(4):690-9.
  8. Brett M, Johnsrude IS, Owen AM (2002): The problem of functional localization in the human brain. Nat Rev Neurosci. Mar;3(3):243-9.

1

Schmidt, MES, supplementary

Attachment 1

Table 1: SNRs for PC1 by Smoothing Kernel Size

PC1 Sigmas (n = 12, 8mm) / 1002-MF / 1003-DV / 1004-DE / 1005-TV / 1006-JA / 1007-YS / 1008-LG / 1009-DVL / 1010-DW / 1011-IP / 1012-RJ / 2001-BD
PLA / 1.8058 / -1.0669 / 1.1645 / -0.0444 / 0.5101 / -0.6768 / -0.0335 / 0.2573 / 0.9653 / -0.9033 / 0.4077 / -2.8253
30 mg / -0.5351 / 3.2219 / 1.0788 / -0.5048 / -0.6185 / -1.8327 / 1.0282 / 0.3049 / -0.3133 / -0.8325 / -0.4707 / 0.2060
200 mg / -1.2049 / -1.3037 / -1.1178 / 0.4102 / -0.1694 / 1.9439 / -0.5543 / -0.4716 / -0.5823 / 1.4315 / 0.2110 / 1.7598
PC1 Sigmas (n = 12, 12mm) / 1002-MF / 1003-DV / 1004-DE / 1005-TV / 1006-JA / 1007-YS / 1008-LG / 1009-DVL / 1010-DW / 1011-IP / 1012-RJ / 2001-BD
PLA / 1.7510 / -1.0083 / 1.2424 / -0.0443 / 0.4316 / -0.5771 / -0.1225 / 0.1702 / 1.0856 / -0.9767 / 0.4333 / -2.7927
30 mg / -0.6186 / 3.0018 / 1.1479 / -0.3611 / -0.6039 / -2.0100 / 1.0465 / 0.3577 / -0.3698 / -0.8047 / -0.3365 / 0.1032
200 mg / -1.1710 / -1.2846 / -1.0784 / 0.2783 / -0.1778 / 2.0233 / -0.4668 / -0.4499 / -0.6345 / 1.4522 / 0.1113 / 1.8163
PC1 Sigmas (n = 12, 16mm) / 1002-MF / 1003-DV / 1004-DE / 1005-TV / 1006-JA / 1007-YS / 1008-LG / 1009-DVL / 1010-DW / 1011-IP / 1012-RJ / 2001-BD
PLA / 1.7087 / -0.9256 / 1.3066 / -0.0224 / 0.3493 / -0.5000 / -0.2228 / 0.0824 / 1.1748 / -1.0485 / 0.4769 / -2.7460
30 mg / -0.7179 / 2.8097 / 1.2047 / -0.2363 / -0.5542 / -2.1292 / 1.0454 / 0.4334 / -0.4064 / -0.7948 / -0.2305 / -0.0189
200 mg / -1.1413 / -1.2416 / -1.0582 / 0.1457 / -0.1927 / 2.0889 / -0.3688 / -0.4306 / -0.6681 / 1.4657 / 0.0090 / 1.8736

1

Schmidt, MES, supplementary