Multimodal magnetic resonance imaging: The coordinated use of multiple, mutually informative probes to understand brain structure and function

Abstract

Differing imaging modalities provide unique channels of information to probe differing aspects of the brain's structural or functional organization. In combination, differing modalities provide complementary and mutually informative data about tissue organization that is more than their sum. We acquired and spatially coregistered data in four MRI modalities—anatomical MRI, functional MRI, diffusion tensor imaging (DTI), and magnetic resonance spectroscopy (MRS)—from 20 healthy adults to understand how interindividual variability in measures from one modality account for variability in measures from other modalities at each voxel of the brain. We detected significant correlations of local volumes with the magnitude of functional activation, suggesting that underlying variation in local volumes contributes to individual variability in functional activation. We also detected significant inverse correlations of NAA (a putative measure of neuronal density and viability) with volumes of white matter in the frontal cortex, with DTI‐based measures of tissue organization within the superior longitudinal fasciculus, and with the magnitude of functional activation and default‐mode activity during simple visual and motor tasks, indicating that substantial variance in local volumes, white matter organization, and functional activation derives from an underlying variability in the number or density of neurons in those regions. Many of these imaging measures correlated with measures of intellectual ability within differing brain tissues and differing neural systems, demonstrating that the neural determinants of intellectual capacity involve numerous and disparate features of brain tissue organization, a conclusion that could be made with confidence only when imaging the same individuals with multiple MRI modalities. Hum Brain Mapp, 2013. © 2011 Wiley Periodicals, Inc.

INTRODUCTION

Magnetic resonance imaging (MRI) uses ingeniously designed combinations of pulsing radiofrequency signals and changing magnetic fields to measure tissue characteristics and their spatial locations in the brain. The energy in the radiofrequency signals is absorbed by nuclei within the atoms and molecules that compose various brain tissues. Those tissues then emit that absorbed energy as a second radiofrequency signal, but now imprinted with the molecular and chemical signatures of the emitting tissues. Because these tissue characteristics are numerically encoded and displayed in the form of an image, MRI is too often regarded as producing a “mere picture,” when in fact the intimate interaction of the radiofrequency signal with chemical nuclei is more accurately regarded as a noninvasive probe of brain tissue. The specific information about the brain that this probe provides depends on the pulse sequences, or the specific combinations of radiofrequency signals and changing magnetic fields, that are used to interrogate the tissue. Some pulse sequences, used in anatomical MRI, provide information about the anatomical organization of gray matter, white matter, and cerebrospinal fluid in the brain. Other sequences, used in functional MRI (fMRI), provide information about time‐varying levels of deoxyhemoglobin in the tissue, which can then be used to identify tissues that change in neural activity in response to performance of a behavioral task. Others, used in diffusion tensor imaging (DTI), provide information about the constraints on the diffusion of water in the brain, which is largely determined by the varying concentrations of cell membranes, organelles, and myelin in white and gray matter. Finally, sequences used in magnetic resonance spectroscopy (MRS) provide information about the concentrations of certain molecules in the brain, including one metabolite, N‐acetyl aspartate (NAA), that is thought to index the density and viability of neurons in the brain [Arnold et al.,2001; Baslow,2003; Edden et al.,2007] and to contribute to signaling between neurons and glia [Lebon et al.,2002]. Although the biological functions of NAA are the subject of controversy, for the sake of simplicity and along with others [Moffett et al.,2006], we will refer to NAA concentrations as indexing neuronal density.

Individually, each of these probes, or MRI modalities, provides a unique channel of information to view and understand one aspect of the brain's structural or functional organization. In combination, however, these modalities provide complementary and mutually informative data about tissue organization that is more than their sum [Filippi,2009]. Moreover, their use in combination can help us understand causal mechanisms between modalities. For example, their combined use can help to determine whether increases or decreases in functional activation detected with fMRI are related to increases or decreases in underlying volume of cortical gray matter in that region and therefore whether abnormal activation is a consequence of, or a compensation for, an underlying anatomical abnormality. DTI, alternatively, can help determine whether anatomical or functional disturbances in two distinct brain regions are associated with, and possibly caused by, underlying disturbances in the anatomical connectivity between those two regions. MRS can help to determine whether regional differences in volume or activation are likely associated with disturbances in the health or number of neurons in that region. Incorporation of multiple, informative imaging modalities therefore can tell us much more about the neural basis of behavior than any single imaging modality can alone. Furthermore, multimodal imaging can aid study of the neurobiological determinants of disease states. The findings from one modality, for example, can help to constrain the interpretations of findings from another modality, thereby improving the neurobiological validity of those findings and interpretations.

We are unaware of any previous studies that have acquired and correlated imaging measures across all four MRI modalities. When using two MRI modalities, voxel‐wise analyses most often have been used, with rare exceptions in preliminary studies [Eichler et al.,2002; Gore et al.,2006; Hayasaka et al.,2006; Irwan et al.,2005; Krakow et al.,1999, Pell et al.,2008], to segment the brain into differing tissues [Devlin et al.,2006; Kabir et al.,2007] or to examine correlations across modalities but within only one or several regions of interest [Barkovich et al.,2006; Nitkunan et al.,2008; Olesen et al.,2003; Toosy et al.,2004].

We report the acquisition, processing, and techniques used for the coregistration of anatomical MRI, fMRI, DTI, and MRS data across the brains of 20 healthy adults. Correlation analyses across modalities were performed at each voxel of the brain to provide preliminary indications of the ways in which interindividual variability in measures from one modality account for variability in measures from other modalities. The imaging measures were also correlated with measures of estimated intelligence and attention to determine which features of brain tissue are associated with higher‐order cognitive functions. Based on our assumptions that cellular composition determines anatomical structure and connectivity in the brain, and that structure in turn determines function, we hypothesized that we would detect significant correlations of neuronal density with measures of local brain volumes and structural connectivity, correlations of local brain volumes with measures of brain function, correlations of neuronal density with functional measures, correlations of structural connectivity with functional measures, and correlations of local brain volume with structural connectivity.

METHODS

We acquired data in four MRI modalities (anatomical MRI, DTI, fMRI, and MRS) from 20 healthy adults and computed voxel‐wise correlations across the various MRI measures to study the interrelations among neuronal density, diffusion anisotropy, fMRI activity, and indices of regional volumes.

Participants

Imaging data were acquired from 20 right‐handed healthy adults (10 males and 10 females, ages from 19 to 45 years, mean age 29.7 ± 7.7 years). The absence of neuropsychiatric illness was assured through administration of the Schedule for Affective Disorders and Schizophrenia [Endicott and Spitzer,1978]. Written informed consent was obtained from each participant.

Cognitive Measures

Intelligence Quotient (IQ) was estimated using the Wechsler Abbreviated Scale of Intelligence [Wechsler,1981]. We used the Connors Continuous Performance Test CPT II (available at: http://www.devdis.com/index.html) to measure attentional capacity and impulse control [Conners,1994]. Participants were requested to press a button when any letter appeared, except for the letter “X”, thereby priming the motor system to respond incorrectly in the minority (10%) of trials when the X appeared, during which time the participant should have inhibited the prepotent inclination to respond. The task effectively identifies errors of impulse control, counting both errors of omission (failing to respond to the target letter) and commission (responding to the nontarget letter “X”). A measure of attentiveness was calculated to indicate how well participants discriminated between targets and nontargets. Raw scores were transformed into T‐scores for use in correlation analyses. Because the Commission and Attentiveness T‐scores were highly intercorrelated (r = 0.91, P < 0.001), we report only results using the Attentiveness measure. Lower T‐scores indicated better performance. IQ and CPT attention measures were not significantly intercorrelated in this sample (Pearson's r < 0.04, P = 0.87).

MRI Acquisition

We collected anatomical T1‐weighted images, diffusion‐weighted (DW) images, functional images, and Multiplanar Chemical Shift Imaging (MPCSI) data in a single scanning session from all individuals in this study. Images were acquired on a GE Signa 3T whole body scanner (Milwaukee, WI) equipped with a body transmitter coil and an 8‐channel head receiver coil. Anatomical MRI, DTI, and fMRI data were acquired with the Array Spatial Sensitivity Encoding Technique (ASSET), a GE version of the parallel imaging technique.

Anatomical MRI

High‐resolution, T1‐weighted images of the brain in sagittal orientation were acquired using fast spoiled gradient recall (FSPGR) sequence: inversion time (TI) = 500 ms, repetition time (TR) = 4.7 ms, echo time (TE) = 1.3 ms, field of view (FOV) = 24 cm, image matrix = 256 × 256, acceleration factor = 2, number of slices = 160, slice thickness = 1 mm.

Diffusion Tensor Imaging

DTI slices were acquired in an axial oblique orientation parallel to the AC‐PC line using single‐shot echo‐planar DTI imaging sequence, with TR = 15,700 ms, TE ∼ 74 ms, FOV = 24 cm, flip = 90°, acquisition matrix = 128 × 128 (acceleration factor = 2) zero‐padded to 256 × 256, slices = 60, slices thickness = 2.5 mm. We acquired three baseline images with b = 0 s/mm², and 25 diffusion weighted images at b = 1,000 s/mm² with diffusion gradients applied in 25 directions sampling three‐dimensional space uniformly [Jones et al.,1999].

Multiplanar Chemical Shift Imaging (MPCSI)

Multiplanar chemical shift imaging (MPCSI) data were prescribed using localizer images acquired with TR = 300 ms, TE = 10 ms, FOV = 24 cm, slice thickness = 10.0 mm, spacing = 2.0 mm, acquisition matrix = 256 × 128, image zero‐padded to 256 × 256. The spectral data were acquired using six axial oblique slices positioned parallel to the AC‐PC line, with the second bottom‐most slice containing the AC‐PC line. Parameters for the MPCSI sequence [Duyn et al.,1993] were: TR = 2,800 ms, TE = 144 ms, spectral width = 2,000 Hz, number of complex data points = 512, FOV = 24 × 24 cm², slice thickness = 10.0 mm, spacing = 2.0 mm, number of phase encoding steps = 24 × 24. Water suppression was achieved using the CHESS sequence. Lipid signal was suppressed by placing eight angulated saturation bands around the brain.

Functional Magnetic Resonance Imaging (fMRI)

T2*‐weighted images were acquired in axial‐oblique slices positioned parallel to the AC‐PC line using a gradient‐recalled single‐shot echo‐planar pulse sequence with TR = 2,200 ms, TE = 30 ms, flip angle = 90°, image matrix = 64 × 64, FOV = 24 × 24 cm, in‐plane resolution = 3.75 × 3.75 mm, number of slices = 34, slice thickness = 3.5 mm, number of imaging volumes = 128. A full set of these images was acquired as participants performed each of three tasks to activate primary motor, visual, and auditory cortices. The order of these tasks was randomized among the 20 subjects.

Visual task

The visual stimulus was a checkerboard flashing at a frequency of 60 Hz and simultaneously rotated at a frequency of 1 Hz [Smith et al.,1998]. The concentric radial checkerboard pattern filled half the display window (the other half of the window was blank). The center of the checkerboard was positioned at the center of the window. The pattern consisted of 60 images per full circle, with each image rotating 6 degrees clockwise from the previous image. The stimulus was programmed and delivered using E‐Prime software (available at: http://www.pstnet.com/products/e-prime) and presented through fiber‐optic goggles. Visual stimulation was presented over seven on‐off cycles in 20‐s blocks of flashing checkerboard that alternated with 20‐s blocks of gaze fixation on a white cross‐hair placed in the middle of a black background.

Motor task

The motor task was identical in timing and similar in design to that of the visual task. The 20‐s blocks of active task consisted of tapping the right index finger as quickly as possible in response to a cue (a symbol changing from “+” to “X”) displayed on the screen. These alternated over seven on‐off cycles with 20‐s epochs of gaze fixation on a white cross‐hair located in the middle of a black background.

Auditory task

Similarly, an auditory stimulus was presented in 20‐s blocks alternating over seven on‐off cycles with 20‐s blocks of gaze fixation on a white cross‐hair located in the middle of a black background. The auditory stimulus consisted of an “alarm whistle” that played at a frequency of 11.025 kHz, delivered using E‐Prime software.

Spatial Normalization of Multimodal Images

Voxel‐wise correlations require the precise normalization of data from various modalities into a template space. Spatially normalizing images from various MRI modalities into one coordinate space is challenging, however, because of differing pixel resolution, image intensity, contrast, and amount of spatial distortion across images. Anatomical MR images usually have the highest resolution, with voxel sizes of 1 × 1 × 1 mm³. DTI and fMRI images usually have lower spatial resolution, with typical voxel sizes of 1 × 1 × 2.5 mm³ and 3 × 3 × 3 mm³, respectively. MRSI images usually have the lowest resolution, with voxel sizes of 10 × 10 × 10 mm³ and with a 2‐mm skip between slices. MRSI data are also typically acquired in only a limited number of slices and therefore do not cover the entire brain.

Spatial Normalization—Overview

Multimodal imaging data were spatially normalized into the coordinate space defined by the high‐resolution, T1‐weighted anatomical image of the template brain. To facilitate voxel‐wise statistical analyses across these images with differing resolutions, we normalized each of the imaging datasets into the common coordinate space by reslicing the DT and fMRI images to voxels 1 × 1 × 1 mm³ in size (for DTI datasets, for example, we first resliced each DW image to 1 × 1 × 1 mm³ voxels using trilinear interpolation and then reconstructed the DTs at this smaller voxel size). fMRI data were smoothed before reslicing, whereas anatomical, DTI, and MPCSI data were not. We then spatially normalized images by first co‐registering them using a rigid body similarity transformation, and then nonlinearly warping the images using a method based on fluid dynamics [Christensen et al.,1994]. However, unique characteristics of the data in each modality required a unique approach to normalization of each modality's data (Fig. 1a).

Figure 1.

Spatially normalizing multimodal datasets. (a) The normalization flowchart. First, each participant's T1‐weighted image (“Source”) was normalized to the template brain, yielding deformation fields D4 and D5. Datasets from other modalities were then normalized independently to the T1‐weighted image of the source brain that had been normalized to the template space. Finally, deformation fields D4 and D5 were used to normalize these datasets into the template space. Hi‐Res T1 = high‐resolution T1‐weighted anatomical MR images; DWI = diffusion weighted imaging datasets; DTI = diffusion tensor imaging datasets; FA = fractional anisotropy; functional (FMRI) = functional imaging datasets; MPCSI = multiplanar chemical shift imaging; Low‐Res T1 = low‐resolution T1‐weighted overlay for MPCSI; Def Field = deformation field; D1,…,D5 = deformation fields estimated in different steps; Compose D4(D5) = single deformation field obtained by composing fields D4 and D5. (b) Examples: spatial normalization of T1‐weighted anatomical images and FA maps to the template Brain. (i) Template image. (ii) Single participant T1‐weighted anatomical “source” image. (iii) Source image normalized to the template brain. (iv) Group‐averaged co‐registered T1‐weighted image of 20 participants in template space. (v) Single participant's source FA map. (vi) Spatially normalized FA map in template space. (vii) Group‐averaged co‐registered FA map of 20 participants in template space. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

Spatial Normalization of Anatomical Images

We registered the T1‐weighted image from each participant (termed the “source” image) to the T1‐weighted template image as follows: (1) We resampled the image to voxel dimensions of 1 × 1 × 1 mm³. (2) We then coregistered the resampled image to the template image using an affine transformation with 12 parameters, such that the mutual information across images was maximized. This step generated a deformation field (D4) between the two images that encoded the transformation mapping each point in the source image into a point in the target image. (3) The coregistered image was then nonlinearly warped into the coordinate space of the template image using fluid dynamics [Bansal et al.,2005; Christensen et al.,1994] while maximizing the mutual information across the images [Maes et al.,1997] to generate a high‐dimensional, nonlinear deformation field (D5).

Spatial Normalization of DTI Data

We first warped the FA image from each participant to the same person's high‐resolution T1‐weighted anatomical image. We used the FA image for this within‐subject warping because its contrast, unlike that of DW images, was similar to that of the T1‐weighted image. We applied two successive deformation fields for this warping, the first (D1) an affine transformation (three rotations, three translations, and global scale) and the second (D2) a high‐dimensional, nonlinear warping of the images based on the dynamics of fluid flow [Christensen et al.,1994]. The application of D1 and D2 transformed the source FA maps in their native space into the coordinate space of the anatomical image for each participant. We then successively applied the same two deformation fields, D4 and D5, previously determined for the coregistration of that participant's anatomical image with the template brain (above). D1, D2, D4, and D5 were then concatenated and applied to the individual DTI datasets to bring them into the template space of high resolution anatomical data, using our seamless procrustean algorithm developed in‐house, for preserving the unique diffusion characteristics encoded in DTI data. The seamless procrustean algorithm estimates the rotation matrix that reorients the tensor at a voxel in template space through the procrustean fit, based on the probability distribution of the fiber orientation at the voxel, which is estimated from the DTI measurements around the voxel [Xu et al.,2003,2008].

Spatial Normalization of MPCSI Data

We normalized MPCSI data using the low‐resolution T1‐weighted overlay image—i.e., the localizer image that was acquired in nearly ideal alignment with the MPCSI data. Because the overlay image had a higher in‐plane resolution than that of the MPCSI data, it coregistered precisely the MPCSI data to the high‐resolution anatomical image of the same person, generating an rigid‐body transformation (D3). We then successively applied deformation fields D4 and D5, determined previously for the coregistration of that participant's anatomical image with the template brain, to the MRSI data to bring them into the template space using trilinear interpolation.

Spatial Normalization of fMRI Data

We first motion‐corrected all functional images of each participant using a rigid transformation to the middle (i.e., the 65th out of 128) functional images of the same participant. We spatially smoothed the motion‐corrected functional images using a Gaussian‐kernel smoothing function with full width at half maximum (FWHM) equal to 8 mm and then we trilinearly interpolated the smoothed images to a resolution of 1 × 1 × 1 mm³, and then coregistered these resampled functional images to the high‐resolution anatomical image of the same person. Finally, we normalized the coregistered functional images to the template brain by successively applying the two deformation fields, D4 and D5, determined previously for the coregistration of that participant's anatomical image with the template brain.

Choice of Template Brain

The coregistrations and the voxel‐wise correlations between imaging datasets conceivably could have depended on the choice of the participant whose brain is designated as the template. We therefore followed a two‐step procedure to select a template brain that was most representative of our cohort by analyzing the morphology of the brain surface defined with high precision in the high‐contrast, high‐resolution T1‐weighted images. First, we identified a brain that was as representative as possible of the demographics of the sample, i.e., demographical data were closest to the mean as possible in terms of age, weight, height, etc. The source brains for all remaining participants were coregistered to this preliminary template. The point correspondences on the surfaces of their cortices were determined, and we computed the distances of the corresponding points on the cerebral surface of the other participants from the surface of the template brain. Then the brain for which all points across the surface are closest (in terms of least squares) to the average of the computed distances was selected as the final template brain for multimodal data analysis. All brains then underwent a second coregistration, this time to this most representative template. We used a single representative brain as a template rather than an averaged brain because a single brain has well‐defined tissue interfaces, such as the CSF‐gray matter or gray‐white matter interfaces.

Confirmation of Registration Accuracy

We assessed the accuracy of our methods for image normalization by visually comparing the warped source image for each participant with the image for the template and with the average image computed from normalized images from all participants. Our visual checks assessed the precise matching of anatomical landmarks, including central sulcus, interhemispheric fissure, anterior commissure, posterior commissure, and genu and splenium of the corpus callosum.

Processing of MRI Data

Data from each MRI modality were processed to correct for spatial distortions, intensity inhomogeneities, motion artifacts, and differing image resolution across imaging modalities. In addition, NAA values were calibrated across participants.

Anatomical Data Processing

We used the computerized algorithm “N3” [Sled et al.,1998] to correct inhomogeneities in pixel intensity across the image that were caused by nonuniformities in the Radio Frequency (RF) field. This algorithm estimates iteratively both the multiplicative bias field and the distribution of the true tissue intensities to eliminate the dependence of the field estimate on anatomy. Extracerebral tissues were removed using an automated tool for extracting the brain [Shattuck and Leahy,2002] combined with manual editing. Connecting dura was removed manually on each slice in sagittal, coronal, and axial views. Finally, the brainstem was transected at the pontomedullary junction. We resliced all images into a standard orientation using anterior‐ and posterior‐commissure landmarks [Talairach and Tournoux,1988] to remove residual head flexion/extension and standard midline landmarks to remove head rotation and tilt. We have found that standardizing orientation improves the accuracy of coregistration to the template brain.

DTI Data Processing

We visually inspected all diffusion weighted images (DWIs) and discarded those having motion larger than 2 to 3 mm or susceptibility artifacts. We corrected eddy‐current spatial distortions along the phase‐encoding direction [Haselgrove and Moore,1996]. We computed the diffusion tensor at each voxel by fitting an ellipsoid to the DWI data acquired along 25 gradient directions and three baseline images [Xu et al.,2007]. To ensure that a tensor D was positive definite, we first decomposed it into the product D = A × A ^T, estimated the matrix A, and computed the tensor D from the product A × A ^T. We reconstructed tensors using DWIs in the original space. Then for each tensor we calculated its fractional anisotropy (FA) value, which expresses the degree of directionality of water diffusion, for use in spatial normalization of the DWI dataset to the template brain.

MPCSI Data Processing

MPCSI data were preprocessed using the software 3DiCSI (available at: http://hatch.cpmc.columbia.edu/software.html) operating on a SUN workstation. The data were spatially filtered using a Hamming window before 2D Fourier transformation. Residual water signal was suppressed using a high pass filter. The data were subjected to 4 Hz Gaussian line broadening and a 1D Fourier Transform to transfer them to the frequency domain. We then selected an ROI in each slice and saved the spectral data of the ROI to a file for further analysis using home‐built software. We used model‐based spectral fitting to model the spectrum in each voxel with a curve, identifying peaks in the spectrum for NAA, Cr, Cho, and lipid metabolites. The area under the peaks for each of the three metabolites was obtained by integrating the corresponding lines (see Fig. 2). To account for variations in receiver gain (RG), we conducted a series of phantom scans with increasing RGs and calculated the ratio of peak area to the noise level for each RG. We used these ratios as correction factors to compensate for the effects of varying RGs on peak areas. Effects of transmitter gains (TG) were also corrected against an arbitrary value of TG₀ = 15.6 dB according to S = S ₀ 10${}^{\text{(TG‐TG}{}_{\text{0}}\text{)/20}}$ [Soher et al.,1996]. Background noise was calculated as the standard deviation of the real part of the complex data in the regions free of signal from metabolites. The average SNR of NAA, defined as the peak height of NAA to the standard deviation of data in the signal‐free region of the spectrum, was greater than 120, an excellent SNR attributable to use of the multichannel coil. We normalized the peak areas by the noise level and used these normalized areas to reconstruct spectroscopic images (SIs) of the metabolites.

Figure 2.

Examples of ¹H MRS spectra. Left, spectra from white matter. Right, spectra from the periphery of the brain showing the peak of residual lipid. The blue curve represents the measured spectra, whereas the other colored curves represent fitted spectra, and the black curve shows the residual spectra. The lipid signal is well separated from the NAA peak, supporting the validity of our NAA measures throughout the brain, including at its periphery. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

In addition to these standard MPCSI data processing steps, we performed partial volume correction on the NAA values. Partial volume effects are pronounced in MRSI data and would limit the accuracy of correlations with data in other modalities. Two sources contribute to MRSI partial volume effects. A large spectroscopic imaging voxel usually consists of varying proportions of gray matter, white matter, and CSF. In addition, the limited number of k‐space sampling points in MRSI produces signal bleeding across voxels, termed a “point‐spread‐function” (PSF) effect, which must be taken into account when correcting for partial volume effects. The PSF of an MRSI signal is a complex function that describes how the MR signal from one voxel spreads to other voxels over the entire field of view. The PSF is determined in part by the k‐space trajectory during data acquisition. It is also determined by the window function for spatial filtering employed prior to Fourier reconstruction to suppress long range signal bleeding but at the price of increasing signal contamination across adjacent voxels.

We calculated the PSF by simulating the MRSI acquisition in an inscribed circle of 24 × 24 grids in k‐space and subsequently spatial filtering the data with a Hamming window function. The resulting 24 × 24 complex array was interpolated to 256 × 256 to match the high resolution MR images. To obtain the compartment images with the same resolution and PSF effect as the MRSI, we segmented the high‐resolution MR images that were coregistered to the MRSI slices into components of gray matter, white matter, and CSF. We then convolved them with the PSF. From these low‐resolution compartment images, we retrieved the NAA concentrations from the gray and white matter in the i‐th voxel using a linear regression model [Pfefferbaum et al.,1999]:

where S _i is the measured data for the metabolite, c _g and c _w are the point spread structural representation of the gray and white matter, M _g and M _w are the gray and white matter contributions to the metabolite signal, and n is noise. As M _g can be expressed by a product of gray to white mater metabolite ratio r _gw and M _w [Pfefferbaum et al.,1999]:

One can find the best gray‐to‐white matter metabolite ratio by first minimizing the error induced by the above regression for the entire data volume and then applying this ratio to locally fit the gray and white matter metabolite contributions using the regression model [Pfefferbaum et al.,1999]. The partial volume‐corrected gray (or white) matter metabolite concentrations at the resolution of the MRS data are then resampled trilinearly into gray (or white) matter metabolite concentrations at a spatial resolution of 1 × 1 × 1 mm³ to permit correlation analyses of MPCSI data at the higher anatomical resolution.

fMRI Data Processing

We processed and analyzed functional images using SPM2 (available at: http://www.fil.ion.ucl.ac.uk/spm/). We first visually inspected images to ensure the absence of motion artifacts. We then used automated procedures in SPM2 to correct motion. All motion estimates were <1 mm displacement and <2 degree rotation along any axis. We removed intensity drifts in the images across time using a sixth order, Butterworth‐type high‐pass filter with a cutoff frequency equal to [3/4] of the task frequency. After the fMRI images were spatially normalized into standard space (above), we computed activations via voxel‐wise General Linear Modeling (GLM) under SPM2, in which the hypothesized hemodynamic response function (HRF) was derived from the task stimulus, which was of an alternating block‐design, with 20 s of stimulus On and 20 s of stimulus Off. The activations were derived from BOLD signal amplitude (i.e., SPM contrast beta images).

Correlation Analyses

We calculated the Pearson's correlation coefficient r to measure the strength of the pair‐wise linear association of two imaging measures at each voxel. These imaging measures included (1) the concentration of NAA (a marker of neuronal density) from MPCSI data, (2) fractional anisotropy (FA, a measure of the directional constraint on the diffusion of water) from DTI data, (3) the BOLD signal amplitude of brain activations (a measure of task‐induced neural responsivity) from fMRI data, and (4) an index of local volume expansion or compression from anatomical MRI data, calculated using volume‐preserved‐warping (VPW) [Xu et al.,2007]. VPW preserves during spatial normalization the intensity weighted volume (i.e., intensity × volume of the voxel) of each voxel. Spatial normalization using VPW condensed relatively larger volumes so that they appeared as voxels of relatively higher signal intensity, and it expanded smaller volumes so that they appeared as voxels of relatively lower signal intensity.

Across all voxels of the brain in template space, we calculated Pearson's correlation coefficient r, and its associated P value [Pagano,2000], for the correlations of (1) NAA with fMRI activation, (2) FA with NAA, (3) local volumes with NAA, and (4) local volumes with fMRI activations. Here we present only a small number of representative slices for all findings. A complete set of slices showing correlations throughout the brain can be found in the Supporting Information.

Our null hypothesis, H ₀, for each analysis was that the correlation between brain measures would equal 0, and the alternate hypothesis, H ₁, was that the correlation would differ significantly from 0. To test these hypotheses, we calculated the test statistic (using n − 2 = 18 degrees of freedom) , and computed its P value. We report voxels identified using a P value threshold <0.05 for the correlation coefficient as significantly correlated, together with the requirement that the correlation at that threshold occurred in a spatial cluster of at least 25 adjacent pixels. All correlation analyses were conducted before application of this conjoint statistical threshold. Based on an approximation formula [Poline et al.,1997], this conjoint requirement yielded an effective P < 0.000005, reducing substantially the false‐positive identification of voxels with significantly correlated imaging measures. This low effective P value enabled us to perform all the combination of tests that we report herein, such that the P values remained much smaller than 0.05 even after considering the multiple statistical comparisons.

RESULTS

Spatially normalized images from each participant matched the template image well. The accuracy of our coregistration algorithm is evident in the image showing the averaged data over all 20 subjects (Fig. 1b), in that the average brain maintains the sharpness of tissue boundaries present in the template image. A noneffective coregistration algorithm would blur these boundaries.

Local Volumes and Functional Activation

We detected statistically significant inverse correlations of VPW values with BOLD signal amplitude of fMRI activation in primary visual cortex (Fig. 3a), indicating that volume compression in this area was associated with a stronger BOLD response.

Figure 3.

Correlation maps of VPW values with visual activation and NAA concentrations, and of NAA concentrations with FA values. Correlation coefficients with P value <0.5 are color encoded and displayed, and those with P value <0.05 are significant correlations. The color bar used to encode correlations is at the bottom right of the figure. (a) Correlation maps of VPW values with visual activation. Activation is in response to stimulation with a flashing checkerboard in 20 healthy participants. Correlation analyses were conducted before application of the fMRI statistical threshold. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Third row: Group‐averaged visual activation BOLD signal amplitude map. The color bar used to encode activations is at the bottom right of the figure. On the right: Scatter‐plots for significantly inverse correlated region within visual cortex (top: r = −0.73, P = −0.0004; bottom: r = −0.57, P = −0.009). (b) Correlations of NAA concentrations with VPW values. The anterior limb of the internal capsule (ALIC) and posterior limb of internal capsule (PLIC) are labeled. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. The extent of NAA data is marked by the white boundary curve. Third row: Group‐averaged VPW values. On the left: Scatter‐plot for an inverse correlation within white matter of the internal capsule (r = −0.68, P = −0.001). On the right: Scatter‐plots for selected significantly inverse correlated regions within prefrontal white matter and white matter (top: r = −0.58, P = −0.007; bottom: r = −0.62, P = −0.003). (c) Correlations of NAA concentrations with FA values. The superior longitudinal fasciculus (SLF) and region of crossing fibers (CF) are labeled. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid on the average T1‐weighted image. The extent of NAA data is marked by the white boundary curve. Third row: Group‐average FA maps. The darker bands within the white matter of the FA maps are areas containing possible crossing fibers (from the superior corona radiate), not gray matter. On the left: Scatter‐plots for selected significantly inverse correlated regions within thalamus and the cortical mantle (top: r = −0.57, P = −0.009; bottom: r = −0.55, P = −0.012). On the right: Scatter‐plots for selected significantly correlated regions within superior longitudinal fasciculus of white matter (top: r = 0.70, P = 0.0006; bottom: r = 0.69, P = 0.0007). [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

Neuronal Density and Local Volumes

We detected statistically significant inverse correlations of NAA with VPW values in white matter regions (Fig. 3b), particularly in the frontal cortex and internal capsule bilaterally, indicating that volume expansion in this area was associated with lower NAA values.

Neuronal Density and White Matter Organization

We detected significant positive correlations of NAA with FA within the superior longitudinal fasciculus of white matter bilaterally (Fig. 3c), indicating that greater neuronal density in these regions was associated with a more constrained diffusion of water and, presumably, with a greater structural organization of white matter tissue. These correlations were not significant in regions where white matter fibers cross one another (regions within white matter that appear darker on the FA map). In contrast, significant inverse associations of NAA with FA were evident in the gray matter of the cortical mantle.

Neuronal Density and Functional Activation

We detected significant positive correlations of NAA with BOLD signal amplitude for fMRI activation in the cuneate and visual association cortices obtained during visual stimulation (Fig. 4a), indicating that greater neuronal density was associated with greater functional activation. Correlations in primary visual cortex, at the center of the peak activation during visual stimulation, were not themselves statistically significant, although the available NAA data extended only into a portion of the area of significant activation because MPCSI data in the posterior cortical mantle were eliminated by placement of the saturation bands needed to suppress lipid signal from the scalp.

Figure 4.

Correlations of NAA concentrations with visual, auditory, and motor activation. Correlation coefficients with P < 0.5 are color encoded and displayed, and those with P < 0.05 are significant correlations. The outermost extent of NAA data is marked by the white boundary curve. The anterior cingulate cortex (ACC) and posterior cingulate cortex (PCC) are labeled. (a) Correlations of NAA concentrations with visual activation during flashing checkerboard task performed by 20 healthy subjects. First row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Second row: Group‐averaged visual activation BOLD signal amplitude map. On the right: Scatter‐plot for significantly correlated region within primary visual cortex (r = 0.68, P = 0.001). (b) Correlations of NAA concentrations with auditory activation. First row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. The “+” sign marks the center of peak activation in auditory cortex. Second row: Group‐averaged auditory activation BOLD signal amplitude map. On the left: Scatter‐plot for a strong positively correlated region within posterior cingulate cortex (top: r = 0.66, P = 0.002; bottom, r = 0.65, P = 0.002). (c) Differences across men and women in correlations of NAA with default‐mode deactivation. The functional activation was generated by auditory stimulation. The statistical model included NAA value and the covariate of sex: Auditory Activation = NAA + sex + NAA × sex. First row: Color‐coded P values associated with the interaction term overlaid on the average T1‐weighted image. The color bar used to encode P values is shown at the top right of the figure. Second row: Group‐averaged auditory activation BOLD signal amplitude map. On the left: Scatter‐plots for NAA × sex interactions in selected significant regions within the anterior cingulate cortex and posterior cingulate cortex (top: P = 0.009; bottom: P = 0.003). (d) Correlations of NAA concentrations with motor activation. First row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Second row: Group‐averaged motor activation BOLD signal amplitude map. On the left: Scatter‐plot for strong positively correlated region within cuneate cortex (top: r = 0.61, P = 0.004) and strong inversely correlated region within primary motor cortex (bottom: r = −0.51, P = −0.022). [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

We also detected positive correlations of NAA with BOLD signal amplitude for fMRI activations in posterior and anterior cingulate cortices during stimulation with the auditory task (Fig. 4b). These were regions that tended to decrease in signal intensity (to deactivate) during auditory stimulation compared with stimulation by only the background scanner noise. In addition, significant differences were detected between men and women in the correlation of fMRI deactivations with NAA values (i.e., a significant activation‐by‐sex interaction) in the anterior and dorsal posterior cingulate cortices during the auditory task. In both regions, NAA correlated positively with the magnitude of deactivation in men but not in women (Fig. 4c).

A similar positive correlation of NAA with BOLD signal amplitude for fMRI deactivation was also detected in the posterior cingulate cortex during motor stimulation (Fig. 4d). Within primary motor cortex itself, however, we detected significant inverse correlations of NAA with fMRI BOLD signal amplitude (Fig. 4d).

Gray Matter Tissue Organization and Functional Activation

FA values in the gray matter of the anterior and posterior cingulate correlated inversely with the magnitude of default‐mode activation during both the auditory and motor tasks (Fig. 5a,b). FA correlated positively with the magnitude of activation of primary visual cortex but inversely with activation of the retrosplenial cortex during the visual task (Fig. 5c).

Figure 5.

Correlations of FA values with auditory, motor, and visual activation. Correlation coefficients with P < 0.5 are color encoded and displayed, and those with P < 0.05 are significant correlations. The anterior cingulate cortex (ACC) and posterior cingulate cortex (PCC) are labeled. (a) Correlations of FA values with auditory activation. First row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Second row: Group‐averaged auditory activation BOLD signal amplitude map. On the left: Scatter‐plot for strong inversely correlated region within primary auditory cortex (top: r = −0.63, P = −0.003; bottom: r = −0.55, P = −0.012). On the right: Scatter‐plot for strong inversely correlated region within anterior cingulate cortex (r = −0.55, P = −0.012). (b) Correlations of FA values with motor activation. First row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Second row: Group‐averaged motor activation BOLD signal amplitude map. On the left: Scatter‐plot for strong inversely correlated region within the anterior cingulate cortex (top: r = −0.67, P = −0.001), posterior cingulate cortex (middle: r = −0.60, P = −0.005), and cuneate cortex (bottom: r = −0.59, P = −0.006). (c) Correlations of FA values with visual activation during flashing checkerboard task performed by 20 healthy subjects. First row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Second row: Group‐averaged visual activation BOLD signal amplitude map. On the left: Scatter‐plots for strong inversely correlated region within primary visual cortex (top: r = −0.54, P = −0.014) and strong positively correlated region within cuneate cortex (bottom: r = 0.74, P = 0.0003). [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

Correlations with Higher Cognitive Functions

The magnitude of default‐mode deactivation during auditory, motor, and visual stimulation correlated inversely with measures of attentional capacity (Fig. 6a,b) and IQ (Fig. 7a,b). Activation of visual association cortices, however, correlated positively with IQ scores (Fig. 7a). In addition, VPW measures correlated inversely with measures of attentional capacity in the anterior and posterior cingulate cortices (Fig. 6c) and positively with IQ scores in inferior prefrontal white matter, the basal ganglia and thalamus, and cortical gray matter throughout the cerebrum (Fig. 7c). FA correlated positively with IQ scores in multiple white matter regions, including the prefrontal cortex, parietal cortex, and cingulate bundle bilaterally (Fig. 7d). Correlations with attentional scores were not affected by covarying for IQ scores, and vice versa. The correlation of VPW values with FA values did not survive correction for multiple comparisons and thus were not included here.

Figure 6.

Correlations of CPT attention T‐score with auditory activation, visual activation, and VPW values. Correlation coefficients with P < 0.5 are color encoded and displayed, and those with P < 0.05 are significant correlations. The color bars used to encode correlations and activations are at the bottom right of the figure. (a) Correlations of CPT attention t‐score with auditory activation. Correlation analyses were conducted before application of the fMRI statistical threshold. First row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Second row: Group‐averaged auditory activation BOLD signal amplitude map. On the right: Scatter‐plot for strong inversely correlated region within primary auditory cortex (top: r = −0.46, P = −0.04) and posterior cingulate cortex (bottom: r = −0.57, P = −0.009). (b) Correlations of CPT attention t‐score with visual activation during flashing checkerboard task performed by 20 healthy subjects. Correlation analyses were conducted before application of the fMRI statistical threshold. First row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Second row: Group‐averaged visual activation BOLD signal amplitude map. On the left: Scatter‐plot for strong inversely correlated region within superior temporal lobe (top: r = −0.51, P = −0.022), and cuneate cortex (bottom: r = −0.79, P = −0.00004). On the right: Scatter‐plot for strong inversely correlated region within superior frontal cortex (top: r = −0.65, P = −0.002), and posterior cingulate cortex (bottom: r = −0.54, P = −0.014). (c) Correlations of CPT attention t‐score with VPW values. First row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Second row: Group‐averaged VPW values. On the right: Scatter‐plot for strong inversely correlated region within anterior cingulate cortex (top: r = −0.58, P = −0.007) and posterior cingulate cortex (bottom: r = −0.62, P = −0.004). [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

Figure 7.

Correlations of full scale IQ with visual and motor activation, VPW values, and FA values. Correlation coefficients with P <0.5 are color encoded and displayed, and those with P < 0.05 are significant correlations. (a) Correlations of full scale IQ with visual activation during flashing checkerboard task performed by 20 healthy subjects. Correlation analyses were conducted before application of the fMRI statistical threshold. First row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Second row: Group‐averaged visual activation BOLD signal amplitude map. On the right: Scatter‐plot for strong inversely correlated region within superior temporal lobe (top: r = −0.56, P = −0.011) and positively correlated region within occipital cortex (top: r = 0.60, P = 0.005). (b) Correlations of full scale IQ with motor activation. Correlation analyses were conducted before application of the fMRI statistical threshold. First row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Second row: Group‐averaged motor activation BOLD signal amplitude map. On the left: Scatter‐plot for strong inversely correlated region within the anterior cingulate cortex (top: r = −0.67, P = −0.001) and middle occipital gyrus (bottom: r = −0.54, P = −0.014). On the right: Scatter‐plot for strong inversely correlated region within posterior cingulate cortex (r = −0.62, P = −0.003). (c) Correlations of full scale IQ with VPW values. First row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Second row: Group‐averaged VPW map. On the left: Scatter‐plot for strong positively correlated region within left anterior corona radiata (top: r = 0.59, P = 0.006) and putamen (bottom: r = 0.65, P = 0.002). On the right: Scatter‐plot for strong positively correlated region within left superior corona radiata (r = 0.60, P = 0.005). (d) Correlations of full scale IQ with FA values. First row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Second row: Group‐averaged FA map onto which the boundaries of strong positively correlated regions were overlaid for better localization of the significant findings. On the left: Scatter‐plot for strong positively correlated region within right anterior corona radiata (r = 0.64, P = 0.002). On the right: Scatter‐plot for strong positively correlated region within right cingulum (top: r = 0.57, P = 0.009) and left superior longitudinal fasciculus (bottom: r = 0.61, P = 0.004). [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

DISCUSSION

We were able to spatially coregister data from all four MRI modalities to a single template brain and then confirm that interindividual variability in measures from one modality frequently was associated with interindividual variability in measures from another modality. In interpreting significant correlations between measures obtained using differing imaging modalities, we assume that cellular composition determines anatomical structure and connectivity in the brain, and that structure in turn determines function, to infer causal mechanisms that underlie the intercorrelations detected between imaging measures. We note each of the causal mechanisms that we propose for the findings of this study, although based on the principles of contemporary neuroscience, constitutes only one of several possible alternative interpretations.

Local Volumes and fMRI Activation

VPW data in template space quantifies local volume changes when warping an image from native space to the template. The output is a voxel‐wise map of the ratio of the volume of a region in native space to the volume of the corresponding region in the template space. A VPW value >1 indicates that the region in native space was warped to a smaller region in template space and, conversely, a VPW value <1 indicates that a region in native space was warped to a larger region in the template. In contrast, data from other imaging modalities (fMRI, MRS, and DTI) were trilinearly interpolated from native space to the template using the deformation field that mapped the native coordinates to those of the template. Therefore, the values for data in these other modalities were identical in native and template space.

We detected significant inverse correlations of VPW measures with functional activation in the primary visual cortex (Fig. 3a), with smaller values in template space, or larger volumes in native space, accompanying more functional activation. These findings suggest that visual activation is closely related to individual variability in volume of the visual cortex. Although MRI is unable to discern the microscopic determinants of local volume expansion in native space, neuroanatomical studies in animal models of development suggest strongly that local cortical volumes are determined by lateral expansion of the cortex, and that this lateral expansion is determined by the number of radial columnar units in that cortical region [Rakic,1988]. This determination of local cortical volumes by lateral expansion is particularly strong in primary sensory cortices, where the laminar organization that determines cortical thickness is highly invariant across individuals. We therefore speculate that the number of radial columnar units, the primordial anatomical building blocks within cortical gray matter, is relatively constant per unit tissue volume of primary visual cortex [Rakic,1988]. A larger volume in native space will thereby contain a greater overall number of columnar units which, if each columnar unit generated a similar degree of task‐related activation, would sum to a larger overall magnitude of activation. Our finding that VPW correlated inversely with functional activation has important implications for understanding findings from fMRI studies, as it suggests that interindividual variability in functional activation may derive, at least in part, from differences in underlying anatomical features of the brain. In particular, our findings suggest that persons who activate more strongly in template space may have an underlying expansion of that anatomical region in native space, prior to normalization to the spatial template.

Neuronal Density and Local Volumes

We detected significant inverse correlations of NAA (a putative measure of neuronal density and viability) with local volumes (VPW measures) in white matter regions, particularly in the frontal cortex and internal capsule bilaterally (Fig. 3b). Thus larger volumes were associated with lower neuronal density in these regions. If we assume that axonal densities in frontal regions and the internal capsule are relatively similar across people in template space, then distributing the same number of axons across a larger volume in native imaging space would reduce the density of those axons. These considerations suggest that the number of axons in these regions may be similar in corresponding voxels across individuals, but that the differing anatomy across people distributes those axons differently. This differential volumetric distribution of similarly numbered white matter axons therefore would produce the inverse correlations of neuronal density with local volumes that we observed.

An alternative explanation is that this inverse correlation is an artifact that derives from differing spatial resolutions of the anatomical and MRS datasets. These differing resolutions would produce differing partial volume effects in regions where white matter, gray matter, and cerebrospinal fluid are in close proximity. Persons who have more CSF from a larger lateral ventricle in a corresponding MRS voxel, for example, would have less NAA in that voxel than would someone who has a smaller lateral ventricle in that voxel. Coregistering the corresponding anatomical images to the template brain in that same person with the larger ventricle would require expansion of white matter surrounding the lateral ventricle to match the smaller template ventricle, thereby ensuring that larger VPW values accompany reduced NAA. Correlations of NAA values with VPW measures, however, were mainly inside gray or white matter, with few significant correlations present at the boundary between tissue types. Moreover, we used the most rigorous and sophisticated methods available to correct NAA values for partial volume effects, thereby substantially reducing the likelihood that the observed correlations of NAA with VPW values were an artifact of differing partial volume effects across these imaging modalities.

Neuronal Density and Tissue Organization

We detected positive correlations of neuronal density (NAA) with DTI‐based measures of tissue organization (FA) within the superior longitudinal fasciculus (Fig. 3c), consistent with positive correlations detected in these same regions in two previous preliminary studies that used chemical shift imaging in healthy human volunteers [Eichler et al.,2002; Irwan et al.,2005]. This finding indicates that greater neuronal density is associated with a more constrained diffusion of water and with a greater structural organization of white matter. A greater number of neurons per unit tissue volume, measured as a greater density, would also increase the density of cell membranes, organelles, and myelin, which in turn would proportionally restrict the diffusion of water in that tissue. Correlations were not significant in regions where white matter fibers cross one another, likely because the fibers there were roughly orthogonal to one another. This crossing of fibers would cause the net FA values within each voxel to equal nearly 0 (and therefore to appear dark on the FA map) because of partial volume averaging. The finding of a significant correlation of neuronal density with FA in regions where the directions of fiber pathways are more coherent, however, has important implications for the study of white matter using DTI. The finding shows that a substantial portion of variance in FA measures likely derives from an underlying variability in the number of neurons in those regions, rather than necessarily from some organizational feature intrinsic to neurons that are otherwise erroneously presumed to be similar in number across people.

We also detected significant inverse correlations of neuronal density (NAA) with measures of the degree of tissue organization (FA) in gray matter of the cortical mantle (Fig. 3c). Cortical tissue consists primarily of nerve cell bodies and neuropil, or arborized dendrites, axons, and synapses, which tend not to have directional coherence within a volume of tissue as large as a DTI or MRS voxel. Greater degrees of arborization of axons and dendrites would increase NAA while also reducing directional coherence [Kroenke et al.,2007], thereby producing an inverse association of NAA with FA values. Therefore, imaging studies that compare FA values in cortical gray matter across individuals should consider that lower FA values may in fact derive from greater axonal and dendritic arborization, and not from less neural tissue per se. We cannot exclude the possibility that the inverse correlation of NAA with FA derives instead from an artifact of coregistering imaging modalities that have differing spatial resolutions (analogous to the artifact postulated for the inverse correlation of VPW and NAA measures), although the highly consistent positive correlations of NAA with FA in white matter regions and inverse correlations in cortical gray matter, albeit sometimes at trend levels of statistical significance (Fig. 3c), weighs against this possibility.

Neuronal Density and fMRI Activation

We detected significant positive correlations of neural density (NAA) with the magnitude of functional activation during visual stimulation in the cuneate cortex and visual association cortices in the inferior parietal cortex (Fig. 4a). This correlation suggests that greater neuronal density produces greater functional activation, although this relationship seems not to be generally true throughout the brain, given that we detected the opposite relationship, an inverse correlation, in the primary motor cortex during motor stimulation (Fig. 4d). These findings suggest that sensory association and primary motor cortices are organized in fundamentally different ways, at least in terms of neuronal density and the local cellular and molecular characteristics that drive the BOLD response to sensory or motor stimulation. The differing correlations, however, could also be attributable to the nature of the tasks used to generate the activation maps. The motor task, for example, was an active task in which participants generated finger movements upon command. The visual stimulation task, in contrast, required only passive participation of our subjects. The passive nature of the task may have contributed fundamentally to the positive correlation in the visual task compared with the inverse correlation in the more active motor task. Consistent with this possibility is the correlation of NAA with fMRI signal change in the auditory cortex during auditory activation, which was also generated using a passive task (listening to white noise ramping up and down).

We also detected significant positive correlations of neuronal density (NAA) with the magnitudes of deactivation in the posterior and anterior cingulate cortices during auditory and motor stimulation. Deactivations during the active task relative to an easier control condition in these brain regions has been well documented in previous studies and has been variously termed either “random episodic silent thinking” [Andreasen et al.,1995] or “default‐mode activity” [Gusnard et al.,2001; Marsh et al.,2006; Mason et al.,2007; McKiernan et al.,2006; Simpson et al.,2001]. These positive correlations indicate that less neuronal density contributes to a greater suppression of default‐mode activity during the active task relative to the easier baseline condition. Although the cellular mechanisms that produce or support this correlation are unknown, we speculate that the reduced neuronal density that accompanies greater suppression of default‐mode activity likely represents a proportionally reduced neuropil in the anterior and posterior cingulate cortices. Reduced neuropil, particularly the pruning of synapses and dendrites, is an important feature of normal development that supports improving attentional processing and cognitive control with advancing age during childhood and adolescence [Giedd et al.,1999; Huttenlocher,1990; Sowell et al.,2003; Tau and Peterson,2010]. This interpretation is consistent with our understanding of the inverse correlation of NAA with FA (above), in which we speculated that reduced neuropil likely produced greater tissue organization within these same cortical gray matter regions.

Gray Matter Tissue Organization and Functional Activation

The inverse correlations of FA values with the magnitude of default‐mode activation in anterior and posterior cingulate cortices during the auditory and motor tasks (Fig. 5) suggests that tissue organization of gray matter in these regions was an important determinant of the magnitude of default‐mode activity, with greater tissue organization accompanying more prominent deactivation. We also detected inverse correlations of NAA with FA in cortical gray matter, and positive correlations of NAA with the magnitude of default‐mode deactivations in the anterior and posterior cingulate cortices. Thus less NAA accompanied more default‐mode deactivation; less NAA accompanied greater FA values; and greater FA accompanied more default‐mode deactivation. Therefore we speculate that the inverse correlation of FA with default‐mode deactivation likely was driven by the underlying influences that NAA values (and the neuronal density that these values are thought to represent) had on FA and default‐mode deactivation. Reduced NAA in cortical gray matter regions, as indicated above, likely represented reduced neuropil, which would have increased tissue organization and FA values. Greater degrees of synaptic and dendritic pruning, the likely cause of reduced neuropil, are thought to enhance attentional processing and cognitive control [Giedd et al.,1999; Huttenlocher,1990; Sowell et al.,2003], cognitive capacities that are believed to suppress the mind‐wandering that is present more during the easier baseline task and that is suppressed during times of greater attentional engagement to produce default‐mode deactivation [Mason et al.,2007; McKiernan et al.,2006].

FA also correlated positively with the magnitude of activation in primary visual cortex but inversely with activation in retrosplenial cortex during the visual task (Fig. 5). These findings, together with those in default‐mode cortices above, suggest that FA may correlate inversely with functional activation in association cortices but positively with activation in primary sensory cortices.

Correlations with Higher Cognitive Functions

Higher IQ and better attentional capacity independently correlated strongly with many of our imaging measures, and in differing brain regions. Better cognitive functioning was consistently associated with greater default‐mode deactivation during auditory, motor, and visual stimulation (Figs. 6 and 7). These findings are consistent with the commonly held view that greater deactivation in default‐mode circuits reflects greater suppression of mind‐wandering activity during a task that engages attention relative to an easier baseline condition [Mason et al.,2007; McKiernan et al.,2006] and with theories that intelligence depends heavily on the capacity to control thought and to suppress irrelevant stimuli. Also consistent with this interpretation of default‐mode correlations was our finding that better attentional capacity on the CPT accompanied larger volumes of the anterior and posterior cingulate cortices (Fig. 6c), the regions most consistently identified as components of default‐mode circuits. Alternatively, greater deactivation in default‐mode regions could also represent greater degrees of mind‐wandering during the easier baseline condition, rather than being determined only by the degree of suppression of mind‐wandering activity. In that case, greater attentional and intellectual capacity would accompany a greater capacity for mind‐wandering activity during the easier baseline condition.

Higher IQ scores accompanied larger volumes of inferior prefrontal white matter, larger cortical gray matter volumes throughout the cerebrum, and larger basal ganglia and thalamus volumes (Fig. 7c). These prefrontal cortices and subcortical nuclei have long been regarded as the neural basis for the cognitive and behavioral control that support better performance on intelligence tests. In addition to these gray matter volumes, higher IQ scores accompanied higher measures of tissue organization (larger FA values) in multiple white matter regions, including the prefrontal cortex, parietal cortex, and cingulate bundle bilaterally (Fig. 7d), consistent with findings from prior DTI studies of IQ correlates [Deary et al.,2006; Schmithorst et al.,2005]. FA correlations with measures of white matter organization likely represent the effects that greater myelination has on increasing FA in white matter and on improving the speed of neural transmission in association fibers that connect gray matter regions across long distances. The correlations could also represent the effects of greater neuronal density, which we have shown correlates positively with FA values in these same white matter regions. Alternatively, the positive correlations in more posterior brain regions, where fiber crossings produce lower FA values, could derive from more fiber crossings and therefore less directional coherence in those who have lower IQ scores.

Limitations

Despite our overall success in acquiring, processing, coregistering, and analyzing multimodal MRI data within a single template space and demonstrating the utility of relating data in one modality to data acquired in another modality, this technology, our approach to image analysis, and our findings have several limitations. Perhaps the most prominent is the differing spatial resolutions of the various imaging modalities, which invariably produces differences in partial volume effects across modalities and the possibility that some of the correlations of measures across modalities may be an artifact that derives from partial volume effects, particularly at the interface of differing cerebral tissues. This kind of artifact seems most plausible for those correlations involving NAA, which is the MRI modality in our dataset that has the largest voxel dimensions. In addition, NAA measures were missing in some portions of the cortical mantle because of the placement of saturation bands needed to prevent contamination of NAA values from lipid signals in the scalp.

Another limitation is inherent in the use of subtraction paradigms to generate measures of functional activation in the fMRI modality. Subtraction paradigms isolate cerebral activity that supports a limited cognitive or behavioral process and therefore measure activity in a limited number of brain regions, leaving many other brain regions unexplored. Implementing more recent procedures that measure cerebral perfusion using MRI [Detre and Wang,2002; Floyd et al.,2003] would provide a measure of resting functional activity throughout the entire brain that could be more versatile and more revealing of biological relationships with other imaging measures in correlation analyses across the cerebrum.

Our methods for coregistration relied heavily on the use of sophisticated methods of linear and nonlinear warping of the images to optimize gray scale and surface boundaries of the brain. Matching certain pulse sequence parameters, including receiver bandwidth, echo spacing, number of echoes, and spatial resolution at the time of image acquisition would have provided identical distortions known to affect the DTI and fMRI imaging datasets and would have aided the accuracy of coregistration, at least across those two imaging modalities [Mulkern et al.,2006; Werring et al.,1999]. Acquiring DTI, fMRI, and MRS at spatially identical resolutions as the anatomical images, however, is not yet possible given current constraints on hardware and pulse sequences that would prohibitively increase scan time and reduce signal‐to‐noise ratios in the images to unacceptable levels.

CONCLUSIONS

These various MRI modalities constitute complementary, mutually informative probes of cerebral tissue that, when used together in the same individuals, improve our understanding of the organizational and functional characteristics of the brain. Findings in this sample of healthy participants revealed that local volumes contribute to individual variability in functional activation and that underlying neuronal density in turn contributes to individual variability in local volumes, functional activation, and commonly used measures of white matter organization. We anticipate that these relationships between imaging measures will be altered in disease states [Eichler et al.,2002]. Although interpreting altered correlations between imaging measures will likely prove challenging, accepting that challenge is the only way to begin determining which structural and functional disturbances are primary, or most fundamental, to the disease process, and which are secondary, or derivative, consequences of those primary disturbances.

Supporting information

Additional Supporting Information may be found in the online version of this article.

Figure 1: Correlation Map of VPW Values with Amplitude of BOLD Signal Change During Visual Activation Correlation coefficients with p‐value < 0.5 are color encoded and displayed, and those with p‐value < 0.05 are significant correlations. The color bar used to encode correlations is at the top right of the figure. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Third row: Group‐averaged visual activation BOLD signal amplitude map. The color bar used to encode activations is at the top right of the figure. Figure 2: Correlations of NAA Concentrations with VPW Values Correlation coefficients with p‐value < 0.5 are color encoded and displayed, and those with p‐value < 0.05 are significant correlations. The color bar used to encode correlations is at the top right of the figure. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. The extent of NAA data is marked by the white boundary curve. Third row: Group‐averaged VPW values. (a) Selected significantly inverse correlated region within white matter of the internal capsule. (b) Selected significantly inverse correlated region within prefrontal white matter and white matter. Figure 3: Correlations of NAA Concentrations with FA Values Correlation coefficients with p‐value < 0.5 are color encoded and displayed, and those with p‐value < 0.05 are significant correlations. The color bar used to encode correlations is at the top right of the figure. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. The extent of NAA data is marked by the white boundary curve. Third row: Group‐averaged FA values. (a) Selected significantly inverse correlated region within thalamus. (b) Selected significantly correlated region within superior longitudinal fasciculus of white matter. Figure 4: Correlations of NAA concentrations with Amplitude of BOLD Signal Change During Visual Activation Correlation coefficients with p‐value < 0.5 are color encoded and displayed, and those with p‐value < 0.05 are significant correlations. The color bar used to encode correlations is at the top right of the figure. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. The extent of NAA data is marked by the white boundary curve. Third row: Group‐averaged visual activation BOLD signal amplitude map. The color bar used to encode activations is at the top right of the figure. Figure 5: Correlations of NAA concentrations with Amplitude of BOLD Signal Change During Auditory Activation Correlation coefficients with p‐value < 0.5 are color encoded and displayed, and those with p‐value < 0.05 are significant correlations. The color bar used to encode correlations is at the top right of the figure. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. The extent of NAA data is marked by the white boundary curve. Third row: Group‐averaged auditory activation BOLD signal amplitude map. The color bar used to encode activations is at the top right of the figure. Figure 6: Correlations of NAA concentrations with Amplitude of BOLD Signal Change During Motor Activation Correlation coefficients with p‐value < 0.5 are color encoded and displayed, and those with p‐value < 0.05 are significant correlations. The color bar used to encode correlations is at the top right of the figure. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. The extent of NAA data is marked by the white boundary curve. Third row: Group‐averaged motor activation BOLD signal amplitude map. The color bar used to encode activations is at the top right of the figure. (a) Selected significantly correlated region within cuneate cortex. (b) Selected significantly inverse correlated region within primary motor cortex. Figure 7: Correlations of FA values with Amplitude of BOLD Signal Change During Auditory Activation Correlation coefficients with p‐value < 0.5 are color encoded and displayed, and those with p‐value < 0.05 are significant correlations. The color bar used to encode correlations is at the top right of the figure. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Third row: Group‐averaged auditory activation BOLD signal amplitude map. The color bar used to encode activations is at the top right of the figure. (a) Selected significantly inverse correlated region within primary auditory cortex. (b) Selected significantly inverse correlated region within anterior cingulate cortex. Figure 8: Correlations of FA values with Amplitude of BOLD Signal Change During Motor Activation Correlation coefficients with p‐value < 0.5 are color encoded and displayed, and those with p‐value < 0.05 are significant correlations. The color bar used to encode correlations is at the top right of the figure. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Third row: Group‐averaged motor activation BOLD signal amplitude map. The color bar used to encode activations is at the top right of the figure. Figure 9: Correlations of FA values with Amplitude of BOLD Signal Change During Visual Activation Correlation coefficients with p‐value < 0.5 are color encoded and displayed, and those with p‐value < 0.05 are significant correlations. The color bar used to encode correlations is at the top right of the figure. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Third row: Group‐averaged visual activation BOLD signal amplitude map. The color bar used to encode activations is at the top right of the figure. Figure 10: Correlations of CPT Attention T‐Score with Amplitude of BOLD Signal Change During Auditory Activation Correlation coefficients with p‐value < 0.5 are color encoded and displayed, and those with p‐value < 0.05 are significant correlations. The color bar used to encode correlations is at the top right of the figure. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Third row: Group‐averaged auditory activation BOLD signal amplitude map. The color bar used to encode activations is at the top right of the figure. (a) Selected significantly inverse correlated region within primary auditory cortex. (b) Selected significantly inverse correlated region within posterior cingulate cortex. Figure 11: Correlations of CPT Attention T‐Score with Amplitude of BOLD Signal Change During Visual Activation Correlation coefficients with p‐value < 0.5 are color encoded and displayed, and those with p‐value < 0.05 are significant correlations. The color bar used to encode correlations is at the top right of the figure. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Third row: Group‐averaged visual activation BOLD signal amplitude map. The color bar used to encode activations is at the top right of the figure. (a) Selected significantly inverse correlated region within superior temporal lobe and cuneate cortex. (b) Selected significantly inverse correlated region within superior frontal cortex and posterior cingulate cortex. Figure 12: Correlations of CPT Attention T‐Score with VPW Values Correlation coefficients with p‐value < 0.5 are color encoded and displayed, and those with p‐value < 0.05 are significant correlations. The color bar used to encode correlations is at the top right of the figure. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Third row: Group‐averaged VPW values. Figure 13: Correlations of Full Scale IQ with Amplitude of BOLD Signal Change During Visual Activation Correlation coefficients with p‐value < 0.5 are color encoded and displayed, and those with p‐value < 0.05 are significant correlations. The color bar used to encode correlations is at the top right of the figure. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Third row: Group‐averaged visual activation BOLD signal amplitude map. The color bar used to encode activations is at the top right of the figure. Figure 14: Correlations of Full Scale IQ with Amplitude of BOLD Signal Change During Motor Activation Correlation coefficients with p‐value < 0.5 are color encoded and displayed, and those with p‐value < 0.05 are significant correlations. The color bar used to encode correlations is at the top right of the figure. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Third row: Group‐averaged motor activation BOLD signal amplitude map. The color bar used to encode activations is at the top right of the figure. (a) Selected significantly inverse correlated region within anterior cingulate cortex and middle occipital gyrus. (b) Selected significantly inverse correlated region within posterior cingulate cortex. Figure 15: Correlations of Full Scale IQ with VPW Values Correlation coefficients with p‐value < 0.5 are color encoded and displayed, and those with p‐value < 0.05 are significant correlations. The color bar used to encode correlations is at the top right of the figure. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Third row: Group‐averaged VPW values. (a) Selected significantly correlated region within putamen and left anterior corona radiata. (b) Selected significantly correlated region within left superior corona radiata. Figure 16: Correlations of Full Scale IQ with FA Values Correlation coefficients with p‐value < 0.5 are color encoded and displayed, and those with p‐value < 0.05 are significant correlations. The color bar used to encode correlations is at the top right of the figure. First row: Group‐averaged T1‐weighted image. Second row: Color‐coded correlation coefficients overlaid onto the average T1‐weighted image. Third row: Group‐averaged FA values. (a) Selected significantly correlated region within right anterior corona radiata and left superior longitudinal fasciculus. (b) Selected significantly correlated region within right cingulum.