In all, 19 patients with schizophrenia and 26 healthy volunteers participated in the study. Participants were assessed as detailed in (Egan et al, 2001
). Exclusions based on the history of other axis I disorders (including substance abuse), significant medical history, or excessive movement in the scanner (corrupting a substantial portion of the acquired images) resulted in the final group of 15 patients and 22 controls reported here. All participants had taken part in the Clinical Brain Disorders Branch ‘Sibling Study' (NCT00001486). Five patients with schizophrenia were recruited as outpatients, whereas the rest were recruited from our inpatient ward (NCT00001247). No T2 hyperintensities were observed on the MRI scans of the participants. At the time of study, all patients were treated with second-generation antipsychotics and adjunctive medications.
DTI sessions were conducted using a GE 1.5T Signa scanner (GE, Milwaukee, WI) with an axial single shot echo planar imaging (EPI) sequence (TE 83.7
ms, 80 slices with isotropic voxel size 2 × 2 × 2
mm, NEX=1, 80 × 110 matrix, field of view: 22
cm) with cardiac gating of each individual slice (TR >10
ms). In all, 107 volumes were acquired for each participant (bmax
; see Supplementary Table S1
), with 42 directions acquired with b
. The directions of diffusion-weighted scans were distributed uniformly in space for each b
value acquired. In order to assess the reproducibility of our measures, ten controls repeated one DTI session twice, with an average time between scans of 61±32 (SD) days.
In separate sessions, three-dimensional structural MRI and fMRI scans were acquired. A T1
-weighted spoiled gradient (SPGR) sequence (TR/TE=2400/5
ms, flip angle=45°, with 124 sagittal slices, 256 × 256 matrix, FOV 24
cm, 0.94 × 0.94 × 1.5
voxel size) at 1.5T (GE) was used for the structural acquisition whereas a 3T system (GE) was used for fMRI EPI acquisition with TR/TE=2000/30
ms, flip angle=90°, FOV 24
cm, 64 × 64 matrix, 3.75 × 3.75 × 6
mm voxel size, 24 contiguous slices. The task paradigm used during fMRI was the 2-back working memory task (30 epochs of 0-back alternating with 2-back as previously described in Callicott et al (2003b)
The Freesurfer software package (http://surfer.nmr.mgh.harvard.edu
) was used to generate eight bilateral cortical () and one thalamic region of interest (ROIs) from the T1
-weighted structural images. ROIs were then transformed into DTI space for tractography. Affine registration (Jenkinson and Smith, 2001
) using a normalized mutual information cost function with 12 degrees of freedom was used for each participant to derive transformation matrices from the non-diffusion-weighted image of the DTI series to the structural T1
-weighted image. The inverse of those transformation matrices was then applied to convert the binary cortical ROIs to diffusion space with trilinear interpolation. Cortical ROIs in diffusion space were thresholded at 0.3, re-binarized, and corrected such that there were no overlapping voxels among regions. To ensure that no white matter voxels were included in the cortical ROIs, all voxels with fractional anisotropy >0.2 were excluded from the cortical ROIs. Bilateral thalamic ROIs in diffusion space were thresholded at 0.6 and re-binarized. These thresholds were established after careful review of overlayed masks and registered T1
and non-diffusion-weighted images. We confirmed visually in each individual that the thalamic mask did not grossly exceed the borders of the thalamus after registration, while encompassing the entire T1
-weighted visible structure (the borders of the thalamus are not easily detectable in the non-diffusion-weighted image). For the cortex, a more liberal threshold for binarization was chosen to retain cortical voxels while preventing the extension of the mask into white matter.
Figure 1 Cortical ROIs in one healthy volunteer. The subdivisions of the cortex used in the analysis are shown for the left hemisphere in a lateral (top) and medial (bottom) view with a color bar for reference. Bilateral ROIs were used in the analysis, whereas (more ...)
All tools for DTI processing belonged to the FMRIB Diffusion Toolbox in the FMRIB Software Library (FSL 4.0; http://www.fmrib.ox.ac.uk/fsl/
). The DTI series was first manually reviewed to eliminate artifacts. Volumes where sudden motion had occurred (resulting in slices with low signal) were eliminated. We believe that the rejection of these volumes did not influence our results because all participants had at least seven volumes with b
and 42 directions with b
. Images were corrected for distortion caused by eddy currents and head motion using an affine registration to the first non-diffusion-weighted volume (Jenkinson and Smith, 2001
The brain was extracted and the diffusion parameters were estimated (Behrens et al, 2007
), giving a probability distribution function of two fibers in each voxel of the brain. For each voxel in the thalamic mask, 5000 samples were sent through the connectivity distribution using probabilistic tractography. Probabilistic tractography was run bilaterally as pilot studies showed that reproducibility was somewhat improved over unilateral measures.
Total percent connectivity was calculated as the number of samples from any thalamic voxel reaching the corresponding cortical ROI divided by the total number of samples from all thalamic voxels reaching any cortical ROI. This is a measure of total tractography-defined connectivity from the thalamus to a particular cortical area, independent of where the tract originated from inside the thalamus.
When >25% of the total number of samples originating in a thalamic voxel reached a cortical ROI, that voxel was assigned to a CDR, considered to be connected to the homonymous cortical ROI. The total number of voxels belonging to a CDR divided by the total number of voxels in the thalamus of that individual was the outcome measure. This is a measure of the relative size of the thalamic CDR connected via tractography to a particular cortical area.
The two dependent variables for each cortical ROI were used as within effects in separate analyses of variance, where diagnostic group was the between effect. Two ROIs (medial temporal and orbito-frontal cortices) were dropped from this analysis due to poor reproducibility across two repeated scans of 10 normal controls (Supplementary Methods a
). In these analyses, main effects of the region were of no interest, and will not be reported further, although we focused on main effects of diagnosis or ROI-by
-diagnosis interactions. If the latter was significant, post-hoc t
-tests were run for each ROI. These were not corrected for multiple comparisons as the overall significance protects from type I error. The effect of possible confounding variables on LPFC percent total connections was assessed as detailed in the Supplementary Methods b
. All statistics of this kind were run with Statistica release 7 (StatSoft, Tulsa, OK).
fMRI was analyzed for a subset of 27 individuals (18 healthy subjects and 9 patients with schizophrenia: Supplementary Methods c, Table S3
) during an n-back working memory task. All fMRI data were preprocessed and analyzed with SPM5 software (http://www.fil.ion.ucl.ac.uk/spm
). All fMRI data sets met the criteria for quality control as described previously (Callicott et al, 2003b
). All fMRI data were pre-processed and spatially normalized to the common stereotactic space provided by the Montreal Neurologic Institute (MNI) and analyzed using a general linear model. For each experimental condition, a box car model convolved with the hemodynamic response function at each voxel was modeled. Linear contrasts were computed producing voxel-wise t-statistical parameter maps for 2-back relative to 0-back and entered into second-level analyses.
Two second-level multiple regression models were tested on the contrast images (2-back–0-back), with the subject as a random factor: one to detect voxels with a significant relationship to LPFC connectivity where both groups were assumed to have equal slopes; the other to detect voxels with significantly different slopes between the diagnostic groups. The first model had LPFC connectivity, diagnosis, age, sex, and 2-back performance as covariates. The second model was constructed including the following predictors: two diagnostic group predictors, two LPFC connectivity predictors (one per diagnostic group), age, sex, and n-back accuracy. The contrast between the two LPFC connectivity vectors (adjusted by group means, age, sex, and n-back performance within each group) was then used to assess group differences in the slope of the regression.
Levels of significance to detect associations between LPFC connectivity and BOLD activation were set at p
=0.005 uncorrected for multiple comparisons. Results are reported in a ROI constructed by normalizing to the MNI template and then averaging the bilateral LPFC ROIs used as tractography targets. The resulting image was smoothed with a 3 × 3 × 3
mm FWHM filter mimicking the fMRI resolution and then thresholded at a value of 0.4 for binarization. Corrections for multiple comparisons were performed at the cluster extent level, using the cluster size threshold computed via AlphaSim, a function available in AFNI (http://afni.nimh.nih.gov/
). This program was run for each individual with 1000 iterations, within the ROI described above, and a cluster connectivity radius of 3.1
mm. AlphaSim estimated the rate of false positive voxels within a cluster that are expected given a threshold of p
<0.005 uncorrected. For a specific α
, any cluster size larger than this number can be considered statistically significant after correction for multiple comparisons. For any group analysis, the critical cluster size is calculated averaging this measure across all individuals. In our analysis, for α
=0.05 the average cluster extent threshold was 621
μl (23 voxels).
The relationship between LPFC total percent connectivity and number of correct responses on the 2-back was assessed with Pearson's r.