UCIfMRIscans

Imaging Data included with the (approx.) 28 subject datasets:

Subjects

Subjects were schizophrenic patients recruited as part of a larger study. All subjects were scanned prior to beginning the larger study. The scanning protocol included a sensorimotor scan(SM), three runs of a working memory task (WM), two runs of a continuous performance task (CPT), a high-resolution sagittal T1 scan, a T2 scan, and the usual localization, shimming, and tuning scans.

Data

The data included are: the unprocessed images in Analyze (SPM2) format, and the pre-processed and analyzed images in each run. The sensorimotor run for each subject was analyzed separately; the three WM runs were analyzed in a single session; and the two CPT runs were analyzed in a single session. T1 data both in Analyze (SPM2) format, and cortically segmented using FreeSurfer, are also available.

Timing details for each scan:

Slice prescription: 24 cm FOV, 28 slices, 5 mm thick with no gap; TR = 3s, TE = 40 ms, 90 deg flip angle.

Sensorimotor task: 80 frames. In active blocks. subjects saw a 3Hz reversing checkerboard while a series of auditory tones played, and subjects tapped their thumb to each finger in sequence. Alternating with active blocks were rest blocks in which subjects lay still and saw a fixation cross. The blocks were 15 s long (the first active block began as the third TR was beginning, following a 6 s countdown). This task was one of the pilot tasks for the FIRST BIRN Phase I study (www.nbirn.net).

Working memory task: Each run was 80 frames. Subjects pushed buttons 1 and 2 in response on every trial—to indicate left and right during the arrow trials, and to indicate “target” or “foil” on the probe trials. The three conditions were Arrows, a memory load of 2 items, and a memory load of 5 items, based on Manoach et al. (1999).

In the memory blocks, a warning frame saying “Learn These” for 500 ms preceded the memory set, which was presented all on one screen for 5 s. This was followed by flashing asterisks (500 ms) to indicate the beginning of the probe period. The ten probe trials were presented as single items for 2s each, with a 300 ms fixation cross between each item. The 2 and 5 item blocks were each 30 s long (10 TRs); a 1 s delay was included to make the timing even. The memory sets were different in every block and every run.

The Arrows blocks were 24 s long (8 TRs) and began immediately with the scanning. Each of ten arrows randomly chosen to point to the left or right was presented for 2 s, with a 300 ms fixation cross between each. The order of the blocks in each run was Arrows, 2 items, Arrows, 5 items, Arrows, 5 items, Arrows, 2 items, Arrows.

CPT: Each run was 150 frames. The scan began with a 10-frame fixation cross. The conditions were Arrows, One digit, or Three digits; the order was Fixation, Arrows, One, Three, One, Three, Arrows, Fixation, Three, Arrows, One, Three, One, Arrows, Fixation. Each block was 10 frames long.

In each condition, a stimulus was presented for 40 ms on each trial, with an interstimulus interval of 1960. In the Arrows condition the arrow pointed to the right 9/10 of the time, and to the left one 1/10 of the trials. The subject pushed a button when the arrow pointed to the left.

Similarly in the One digit condition, a single digit was presented with the same timing as the arrows; the zero appeared with a 1/10 frequency, and the subject pushed a button when a zero appeared. In the Three digit condition, three digits appeared simultaneously; the subject was instructed to push the button when a zero appeared in the middle of the screen, which it did on 1/10 of the trials.

Preprocessing

All EPI (fMRI) data were pre-processed and analyzed using SPM2. (A concise overview to the program and methods can be found here: http://www.fil.ion.ucl.ac.uk/spm/software/ and http://www.fil.ion.ucl.ac.uk/spm/doc/intro/ ). The first two images were discarded (note that the timing described above does not take that into account and it must be included in the design matrix determination).

The preprocessing steps included:

1) Motion detection and correction

2) Co-registration and normalization to the SPM EPI template (MNI template)

a. Without B0 mapping, co-registering to a T1 anatomical scan and then to the template can make distortions worse; thus the fMRI data (EPI) was normalized to the canonical template from the same imaging modality.

3) Smoothing with an 8 mm FWHM 3D Gaussian

All algorithms were run with the SPM2 default settings where applicable. No spike detection or ghosting detection was used. No subjects were excluded for head movement.

1. The motion correction algorithm: (from the SPM Help files) This routine realigns a time-series of images acquired from the same subject using a least squares approach and a 6 parameter (rigid body) spatial transformation. For a single task, all runs were realigned using bilinear interpolation to the first image of the first run after discarding any unstable images. No slice-timing or spin-history corrections were made.

2. Co-registration: (from the SPM Help files) Normally the program has two modes of operation: 1) If the modalities of the target image(s) and the object image(s) are the same, then the program performs within mode coregistration by minimising the sum of squares difference between the target and object. 2) If the modalities differ, then the following is performed:

i. Affine normalisation of object to a template of the same modality, and affine normalisation of the target to a template of the same modality. Only the parameters which describe rigid body transformations are allowed to differ between these normalisations. This produces a rough coregistration of the images.

ii. The images are partitioned into gray matter, white matter, csf and (possibly) scalp using spm_segment.m. The mappings from images to templates derived from the previous step are used to map from the images to a set of a-priori probability images of GM, WM and CSF. iii. These partitions are then registered together simultaneously, using the results of step i as a starting estimate.

b. We were co-registering EPI and EPI, so mode 1 was used.

c. A bibliography of relevant articles on these algorithms can be found here: http://www.fil.ion.ucl.ac.uk/spm/doc/biblio/Keyword/REGISTRATION.html

3. Normalization: (from SPM Help) This module spatially (stereotactically) normalizes MRI, PET or SPECT images into a standard space defined by some ideal model or template image[s]. The template images supplied with SPM conform to the space defined by the ICBM, NIH P-20 project, and approximate that of the the space described in the atlas of Talairach and Tournoux (1988). The transformation can also be applied to any other image that has been coregistered with these scans.

a. Mechanism: Generally, the algorithms work by minimising the sum of squares difference between the image which is to be normalised, and a linear combination of one or more template images. For the least squares registration to produce an unbiased estimate of the spatial transformation, the image contrast in the templates (or linear combination of templates) should be similar to that of the image from which the spatial normalization is derived. The registration simply searches for an optimum solution. If the starting estimates are not good, then the optimum it finds may not find the global optimum.

The first step of the normalization is to determine the optimum 12-parameter affine transformation. Initially, the registration is performed by matching the whole of the head (including the scalp) to the template. Following this, the registration proceeded by only matching the brains together, by appropriate weighting of the template voxels. This is a completely automated procedure (that does not require scalp editing) that discounts the confounding effects of skull and scalp differences. A Bayesian framework is used, such that the registration searches for the solution that maximizes the a posteriori probability of it being correct. i.e., it maximizes the product of the likelihood function (derived from the residual squared difference) and the prior function (which is based on the probability of obtaining a particular set of zooms and shears).

The affine registration is followed by estimating nonlinear deformations, whereby the deformations are defined by a linear combination of three dimensional discrete cosine transform (DCT) basis functions. The default options result in each of the deformation fields being described by 1176 parameters, where these represent the coefficients of the deformations in three orthogonal directions. The matching involved simultaneously minimizing the membrane energies of the deformation fields and the residual squared difference between the images and template(s).

b. The images were interpolated using bilinear interpolation, and resliced to 8 x 8 x8 mm voxels (prior to smoothing as noted above).

c. A bibliography of relevant articles on these algorithms can be found here: http://www.fil.ion.ucl.ac.uk/spm/doc/biblio/Keyword/NORMALISATION.html

Analysis:

Each task was analyzed separately; the sensorimotor as a single run, the 3 WM scans together, and the 2 CPT scans together. The design matrix for each modeled the on/off timing of the various conditions as a boxcar, convolved with the canonical hemodynamic response function. Temporal derivatives were included as covariates. The images were scaled globally (see below). Highpass temporal filtering was set as needed for each design. Low-pass temporal filtering used the canonical hemodynamic response function.

Design matrix (from SPM Help): The design matrix defines the experimental design and the nature of hypothesis testing to be implemented. The design matrix has one row for each scan and one column for each effect or explanatory variable. (e.g. regressor or stimulus function). The parameters are estimated in a least squares sense using the general linear model. Specific profiles within these parameters are tested using a linear compound or contrast with the T or F statistic. The resulting statistical map constitutes an SPM. The SPM{T}/{F} is then characterized in terms of focal or regional differences by assuming that (under the null hypothesis) the components of the SPM (i.e. residual fields) behave as smooth stationary Gaussian fields.

Temporal correlations (from SPM Help): Serial correlations in fast fMRI time-series are dealt with as described in spm_spm. At this stage you need to specific the filtering that will be applied to the data (and design matrix). This filtering is important to ensure that bias in estimates of the standard error are minimized. This bias results from a discrepancy between the estimated (or assumed) auto-correlation structure of the data and the actual intrinsic correlations. The intrinsic correlations will be estimated automatically using an AR(1) model during parameter estimation. The discrepancy between estimated and actual intrinsic (i.e. prior to filtering) correlations are greatest at low frequencies. Therefore specification of the high-pass component of the filter is particularly important. High pass filtering is now implemented at the level of the filtering matrix K (as opposed to entering as confounds in the design matrix). The default cutoff period is twice the maximum time interval between the most frequently occurring event or epoch (i.e the minium of all maximum intervals over event or epochs).

Global scaling (from the SPM archives, Thu, 30 Nov 2000, From: Stefan Kiebel <skiebel@FIL.ION.UCL.AC.UK>): In SPM99, a global mean intensity (GMI) is computed for each image. The procedure used by default defines the GMI as

m = mean(Y(Y > mean(Y)/8));

i.e. take the mean of an image Y, divide this value by 8, use it as a threshold for the image and take the mean of all voxels above this threshold. For some reasons, this computation has been found to give you an estimate of the mean of all intracortical voxels in functional images.

For fMRI, SPM (by default) scales each image within a session with 100/(mean of the GMIs of this session). For a single session, this does not have any effect on the parameter estimation or inference, etc. It is just a global scaling of the whole session data. For multiple sessions, this grand mean session scaling has the advantageous effect that each session is scaled to the same grand mean of 100.

Various contrasts were created to examine specific activations. The following table lists major contrast images for each task.

Task Contrast Description

SM con_0003.img Main effect(On)

CPT con_0008.img Main effect(Arrow)

CPT con_0009.img Main effect(Single)

CPT con_0010.img Main effect(Triple)

CPT con_0011.img Differential effect(Arrow<Single)

CPT con_0012.img Differential effect(Arrow<Triple)

CPT con_0013.img Differential effect(Single<Triple)

CPT con_0014.img Differential effect(Arrow<Single +Triple)

WM con_0008.img Main effect(Two)

WM con_0009.img Main effect(Five)

WM con_0010.img Differential effect(Two<Five)