Applying FSL to the FIAC data: Model‐based and model‐free analysis of voice and sentence repetition priming

Abstract

This article presents results obtained from applying various tools from FSL (FMRIB Software Library) to data from the repetition priming experiment used for the HBM'05 Functional Image Analysis Contest. We present analyses from the model‐based General Linear Model (GLM) tool (FEAT) and from the model‐free independent component analysis tool (MELODIC). We also discuss the application of tools for the correction of image distortions prior to the statistical analysis and the utility of recent advances in functional magnetic resonance imaging (FMRI) time series modeling and inference such as the use of optimal constrained HRF basis function modeling and mixture modeling inference. The combination of hemodynamic response function (HRF) and mixture modeling, in particular, revealed that both sentence content and speaker voice priming effects occurred bilaterally along the length of the superior temporal sulcus (STS). These results suggest that both are processed in a single underlying system without any significant asymmetries for content vs. voice processing. Hum Brain Mapp, 2006. © 2006 Wiley‐Liss, Inc.

INTRODUCTION

The techniques available for the interrogation and analysis of neuroimaging data have a large influence in determining the flexibility, sensitivity, and scope of neuroimaging experiments.

The Functional Image Analysis Contest (FIAC), organized as part of the 11th Annual Meeting of the Organization for Human Brain Mapping (Toronto, 2005), was designed to investigate the interactions between the neuroscientific interpretation of results and the particular type of data analysis. In this article we present analysis results based on the tools available in the FMRIB Software Library (FSL).

The aim of the original functional magnetic resonance imaging (FMRI) experiment was to investigate cortical regions sensitive to auditory sentence comprehension as well as the voice of the speaker using a repetition priming paradigm. The experiment manipulated the sentence content (repeated vs. difference sentences) and the speaker's identity (same vs. different speaker) in a 2 × 2 factorial design; participants passively listened to auditory sentences in four conditions [Dehaene‐Lambertz et al., 2006]:

• 1Different sentences, different speakers (DStDSp). This condition served as the baseline for evaluating priming effects.
• 2Same sentence, different speakers (SStDSp) was used to evaluate repetition priming effects due to repeated sentences.
• 3Different sentence, same speakers (DStSSp) was used to evaluate repetition priming effects due to hearing the same voice across sentences.
• 4Same sentence, same speaker (SStSSp). This condition represents the interaction between sentence content and voice.

The experiment was run four times for each subject, twice using a blocked design and twice using a pseudo‐randomized event‐related design. In the blocked design, each 20‐second block of sentences was separated by a 9‐second block of silent rest. There was no comparable “resting” condition in the event‐related design. Previous studies have shown that at a neural level, repetition priming tends to manifest as a reduction in activation in auditory association areas [Dehaene et al., 2001; Vuilleumier et al., 2002], possibly due to habituation of neuronal responses [Desimone, 1996]. Consequently, by manipulating the relation between sentences one can investigate sensitivity to sentence content and the spoken voice.

FMRIB Software Library: Overview

FMRIB's Software Library (FSL) is a comprehensive library of image analysis and statistical tools for FMRI, MRI, and DTI brain imaging data, freely available as both source and binary distributions for various computing platforms. Most of the tools can be run both from the command line and as GUIs (“point‐and‐click” graphical user interfaces). As well as analysis tools, FSL also includes an intuitive yet powerful 3‐D/4‐D image display tool, FSLView, including multiple orthogonal or lightbox views, 3‐D rendering, time series display, image editing, and histogram viewing. For more detail on the FSL analysis tools, see http://www.fmrib.ox.ac.uk/fsl and Smith et al. [2004] (also see Table I for a list of FSL tools).

Table I.

List of FSL tools

FMRIB	Functional Magnetic Resonance Imaging of the Brain Centre, Dept. of Clinical Neurology, University of Oxford, Oxford, UK (see http://www.fmrib.ox.ac.uk).
FSL	FMRIB's Software Library; a collection of tools for the analysis of neuroimaging data.
BET	Brain Extraction Tool; segments brain from nonbrain in structural and functional data, and models skull and scalp surfaces.
PRELUDE	Phase Region Expanding Labeller for Unwrapping Discrete Estimates; performs 3D phase unwrapping of fieldmap images.
FUGUE	FMRIB's Utility for Geometrically Unwarping EPIs; performs unwarping of an EPI image based on unwrapped fieldmap data.
FLIRT	FMRIB's Linear Image Registration Tool; fully automated robust and accurate tool for linear (affine) intra‐ and inter‐modal brain image registration.
MCFLIRT	Motion Correction using FMRIB's Linear Image Registration Tool; intra‐modal motion correction tool.
FEAT	FMRI Expert Analysis Tool; model‐based FMRI analysis tool with a powerful GUI. FEAT provides a single consistent interface for setting up all aspects of preprocessing, linear modeling, registration, and inference.
FILM	FMRIB's Improved Linear Model; tool for estimating first‐level GLMs using pre‐whitening.
FLAME	FMRIB's Local Analysis of Mixed Effects; tool for estimating higher‐level mixed‐effects (between session/subject) GLMs using Bayesian estimation techniques.
FLOBS	FMRIB's Linear Optimised Basis Sets; program for generating optimal basis sets for use in HRF convolution in FMRI linear modelling such as in FEAT.
MELODIC	Multivariate Exploratory Linear and Optimised Decomposition into Independent Components; ICA‐based model‐free analysis of 4‐D Data.

MODEL‐BASED FMRI ANALYSIS USING FEAT

The FMRI Expert Analysis Tool (FEAT) is an advanced General Linear Model (GLM)‐based FMRI analysis tool with a straightforward but powerful GUI, carrying out data preprocessing (including slice timing correction and motion correction); first‐level GLM time‐series analysis with prewhitening; registration to subject‐specific structural images and standard space; and mixed‐effects group analysis using Bayesian estimation techniques.

Complete analysis for a single simple FMRI experiment can often be set up in less than 1 minute, while a more complex experiment typically need take no longer than 5 minutes to set up. Multiple experiments having the same design can be analyzed with a single GUI setup. A complete setup can be saved to file, for easy reloading, amendment and rerunning later, or to be used in script‐based analysis of multiple experiments. The FEAT program produces a web page analysis report (including activation overlay images, activation cluster tables, time‐course plots of data vs. model, registration overlay images and an “Analysis Methods” paragraph describing the exact analysis carried out, including references and parameter value settings).

Data Preprocessing

All 60 FIAC functional datasets (block and event‐related) were preprocessed using the following analysis steps: motion correction using MCFLIRT [Jenkinson et al., 2002]; nonbrain removal using BET [Smith, 2002]; spatial smoothing using a Gaussian kernel of full‐width‐at‐half‐maximum (FWHM) of 5 mm; mean‐based intensity normalization of all volumes by the same factor; high‐pass temporal filtering (Gaussian‐weighted least‐squares straight line fitting, sigma = 50 seconds). All this preprocessing is easily set up in the “Prestats” part of the FEAT GUI (Fig. 1).

Figure 1.

Setup of prestatistical processing steps in the FEAT GUI.

First‐Level General Linear Modeling

The data from the FIAC block and event‐related experiments were supplied with paradigm timing files. These were easily converted into the timing file format required by FEAT, with one timing file per explanatory variable (EV) a column in the design matrix. In total, 8 EVs were used for each design matrix, the four different conditions together with a temporal derivative per primary regressor, to account for potential misspecification of the hemodynamic delay. Figure 2 shows example first‐level design matrices generated by FEAT for one block‐design dataset and one event‐related dataset.

Figure 2.

First‐level design matrices used for the model‐based (GLM) analysis. The columns of the design matrices show graphical representation of the GLM regressors. For the FIAC data there were 4 regressors modeling the BOLD changes during the 4 experimental conditions, together with temporal derivatives (each regressor is immediately followed by its temporal derivative), which are used to account for misspecification of the hemodynamic lag. The design matrix plots also show the width of the temporal high‐pass filter (far left; any low‐frequency fluctuation with periodicity larger than the indicated red bar is removed from the data). At the bottom are the contrast vectors used for testing different primary or differential BOLD changes.

FILM parameter estimation.

Time‐series statistical analysis was carried out using FILM (FMRIB's Improved Linear Model), which corrects each voxel's time‐series for temporal smoothness. Unless this smoothness is correctly accounted for, the time‐series analysis is at best inefficient (in terms of sensitivity to true activation) and at worst statistically invalid (in terms of inflated type1 error due to incorrect degrees‐of‐freedom, DF). To achieve this, FILM uses voxel‐wise autocorrelation estimation involving temporal high‐pass filtering (to remove the worst of the large‐scale, nonstationary components and low‐frequency noise), estimation and regularization of the autocorrelation coefficients in the frequency domain, followed by spatial regularization of the estimates using nonlinear smoothing, within tissue type. This approach has been demonstrated to have close to zero bias at probability levels as low as 0.00001 while obtaining close to optimally efficient estimation of the model parameters, giving greater sensitivity to activation detection [Woolrich et al., 2001].

Contrasts and T/Z statistics.

Each regressor (column) in the design matrix generates a parameter estimate as a result of the model‐fitting. Questions (such as “where is the response to the different‐speaker condition greater than the response to same‐speaker”) are asked by defining “contrasts” of parameter estimates (COPEs), which specify how to take linear combinations of the parameter estimates.

The resulting COPEs can then be tested for statistical significance (against the null‐hypothesis of zero effect size) by converting into T and/or F statistics, taking the estimated error in the fitting into account. For the FIAC data, the primary questions of interest are:

• Primary areas of activation under the four different conditions (SStSSp, SStDSp, DStSSp, DStDSp; contrasts C1–C4).

• The effect of the speaker (under same sentence condition, different sentence condition, and averaged over these; contrasts C5–C7). Of particular interest here is contrast C6, the direct comparison of the DStSSp and DStDSp conditions (voice repetition priming).

• The effect of sentence (under same‐speaker condition, different‐speaker condition and averaged; contrasts C8–C10). Of particular interest here is contrast C9, the direct comparison of the SStDSp and DStDSp conditions (sentence repetition priming).

• Positive (C11) or negative (C12) sentence‐speaker interaction and the overall maximum effect size of the repetition suppression (cognitive interaction: BOLD amplitude decrease during SStSSp when compared to DStDSp; contrast C13).

Registration and correction for geometric EPI distortions.

FEAT carries out preprocessing and time‐series statistics before estimating the alignment transformation between the low‐resolution FMRI data and standard (e.g., MNI152) space. By default, FEAT achieves this using FLIRT [Jenkinson et al., 2002] to first align an example low‐resolution FMRI image to the same subject's high‐resolution T1‐weighted structural image, and then registering the structural image to a standard space average. In Smith et al. [2005], we showed the added value of this two‐stage approach in terms of decreased registration‐induced between‐subject variance, compared with registering directly from FMRI to standard space.

In addition to the linear transforms applied in this way, it is possible to unwarp FMRI EPI (echo‐planar imaging) data prior to registration to the structural image. Distortion and dropout of EPI‐based functional images is a particular problem for high‐field (3T and higher) MR scanners such as the one used for the acquisition of the FIAC data. Such artefacts are particularly noticeable in the inferior temporal and frontal lobes and limit the use of standard FMRI or diffusion imaging techniques in these areas [Devlin et al., 2000; Jezzard and Balaban, 1995].

In addition, the distortions can also lead to global errors in registration, causing errors in the spatial localization of activations from any brain region, including those where there is little or no distortion present. If one acquires B0 fieldmap images one can use these to estimate and correct the EPI distortion.

In the case of the FIAC data the EPI images were acquired using an echo time of 35 ms, which is likely to result in substantial distortions at 3T. However, the FIAC data includes raw fieldmap images, which we used to undistort the FMRI data before running the linear alignments. The fieldmap images were first phase‐unwrapped using PRELUDE (Phase Region Expanding Labeller for Unwrapping Discrete Estimates; Jenkinson [2003]). This corrects for the fact that MR phase measurements are wrapped within the range 0–2π. A corrected fieldmap was calculated from the difference of two unwrapped phase images. This fieldmap was then registered to the FMRI EPI data (using 6 DF), the unwarping was applied to an example EPI image using FUGUE (FMRIB's Utility for Geometrically Unwarping EPIs), and the undistorted EPI image was registered to the subject's structural (using 7 DF). In the registration steps, the signal loss (resulting from through‐slice field gradients) was calculated and used as a cost function mask to exclude voxels in the image where the signal loss was severe [Jenkinson, 2004].

All of the above was carried out using command‐line programs supplied with FSL, and is inserted as a “correction” to the FLIRT‐estimated registrations after the first‐level FEAT analyses have completed, before running cross‐subject FEAT analyses.

Higher‐Level GLM

In the present study the challenge is to address questions about activation effects in generic populations of subjects, i.e., analyse the data in such a way as to allow for hypothesis tests at the population level. In order to be able to generate results that accurately extend to the wider population, we need to account for the fact that these limited samples from the population are random quantities with associated random effects variances on top of the within‐session fixed‐effects variance contribution [Beckmann et al., 2003a; Holmes and Friston, 1998].

In FEAT, such mixed‐effects group statistics are generated using a hierarchical summary statistics approach where the different levels of the hierarchy are separate GLMs. That is, the within‐session first‐level, the within‐subject‐cross‐session second‐level and the cross‐subject third‐level [Beckmann et al., 2003a; Woolrich et al., 2004a]. Note that in many studies, such as those with a single session per subject, only a two‐level approach is needed.

A practical fully Bayesian inference technique is used for higher‐level analyses in FEAT, referred to as FLAME (FMRIB's Local Analysis of Mixed Effects; Woolrich et al. [2004a]). FLAME is a two‐stage process: a fast approach using maximum a posteriori estimates and a slower, more accurate approach using Markov Chain Monte Carlo (MCMC). By taking into account the correct lower‐level summary statistics (estimates of the mean, variance, and DFs), a substantial improvement in higher‐level activation estimation accuracy can be obtained [see Woolrich et al., 2004a, for details].

For the FIAC analyses, each subject's different sessions are combined into a subject‐specific mean at the second level. These means are all estimated as part of one single second‐level design (as opposed to a separate second‐level design for each subject) because the number of sessions per subject is too small to obtain a robust estimate of the between‐session variance separately for each subject; hence, the session variance estimation is pooled across subjects to improve accuracy (at the expense of assuming similar between‐session variance between subjects). At the third (top) level these subject means are then used to calculate mixed‐effects group mean estimates. The higher‐level analyses are run separately for each contrast specified in the first‐level designs. Figure 3 gives a graphical representation of the multilevel hierarchical GLM approach used for the model‐based analysis of the FIAC data.

Figure 3.

Hierarchical multilevel GLM approach for mixed‐effects analysis of the FIAC data: at the first level, each individual session's data is analysed using first‐level designs described in Figure 2. At the intermediate level, each subject's different sessions are combined into mixed‐effects subject‐specific means. At the third (top) level, these subject means are then used to calculate mixed‐effects group mean estimates.

Model‐Based Group Mean Results

The group mean mixed‐effects Z (Gaussianized T/F statistics; Jenkinson and Woolrich [2000]) statistic images were thresholded using clusters determined by Z > 2.3 and a (corrected) cluster significance threshold of P < 0.05 assuming a Gaussian random field for the Z‐statistics [Worsley et al., 1992].

Figure 4 shows the cross‐subject mean for the different‐sentence‐different‐speaker condition (contrast 4), averaged over all 16 subjects (blue: block‐design; red: event‐related design). In the context of this experiment this condition is typically referred to as the “baseline” condition. This condition generates the highest level of activation, i.e., largest estimated BOLD amplitude modulation, and repetition priming in other conditions is expected to cause suppression of the activation level relative to this baseline.

Figure 4.

Cross‐subject mean (contrast 4: different sentences different speakers). Results are shown for the block‐design data (blue) and the event‐related design (red). All 2‐D image slices are shown left‐right reversed. 3‐D rendering carried out using FSLView.

Significant BOLD modulation was observed bilaterally in Heschl's gyrus (HG), planum temporale, planum polare, lateral superior temporal gyrus, superior temporal sulcus (STS), and middle temporal gyrus. These regions encompass primary and secondary auditory cortices as well as extensive auditory association cortex and correspond to previous studies reporting auditory sentence processing [Cardillo et al., 2004; Crinion et al., 2003]. There were no substantial differences in the pattern of activation between the event‐related and blocked versions. For the differential contrasts of interest, Table II lists all significant clusters and the location of the maximum statistics in MNI152 space.

Table II.

FMRI activation cluster inferred from the model‐based GLM analysis

Location	Cluster size(mm3)	Statistical significance, P	Maximum statistics	Coordinates: x, y, z *
Effects of speaker repetition restricted to same sentence condition (Fig. 5, contrast C5)
Right Broca's area	7120	< 0.0001	Z = 3.23	46, 20, 24
Left STS/MTG	6336	< 0.0003	Z = 3.44	−54, −58, 6
Main effects of sentence repetition (Fig. 6, contrast C9)
Left STS/MTG (middle portion)	3592	< 0.007	Z = 3.69	−62, −14, −12
Interaction effects between sentence content and voice (negative interaction C12)
Right Broca's area	3136	< 0.02	Z = 3.33	44, 8, 26
Left posterior STS/MTG	2920	< 0.021	Z = 3.62	−60, −66, 6
Maximum effect of repetition surpression (Fig. 7, contrast C13)
Left STS/MTG	12,904	< 10−6	Z = 4.35	−58, −16, −8
Broca's area	4776	< 0.003	Z = 4.13	−58, 16, 18
Pos. intraparietal sulcus	4528	< 0.004	Z = 3.19	−32, −64, 42
Right ant. STS/MTG	3848	< 0.011	Z = 3.54	62, 0, −12
Main effects of speaker repetition (Fig. 8, FLOBS and mixture modeling on contrasts C6)
bilat. STG extending onto	7544	> 0.95a	F = 4.23	−64, −24, 2
lat. Heschl's Gyrus	6584	> 0.95a	F = 3.65	68, −20, 4

^*Coordinates are reported in mm from the anterior commissure (AC) in MNI152 standard space.

^aAlternative hypothesis test based on nonspatial Gaussian/Gamma mixture modeling of the F‐statistics image.

Priming effects due to speaker repetition.

Compared to the sentence effect, repetition priming due to the repetition of the speaker is expected to elicit a smaller reduction in BOLD amplitude. Indeed, there was no significant main effect of voice repetition (contrast 6: DStSSp < DStDSp) in either the event‐related or blocked versions when using the default null‐hypothesis testing approach with Gaussian Random Field (GRF) theory. A significant cluster was seen for the simple main effect of voice repetition (contrast 5: SStSSp < SStDSp) in the left STS (see Fig. 5; event‐related data only).

Figure 5.

Repetition effect of speakers (contrast 5: SStSSp < SStDSp) estimated from the event‐related datasets (red) when sentence content is not changed between different speakers. No supra‐thresholded voxels were found in the block‐design data.

Priming effects due to sentence repetition.

BOLD amplitude decreases due to the repetition of sentences are explicitly tested by contrasts 8–10. Figure 6 shows the primary sentence effect of interest, the priming effect of different speakers presenting the same sentence (contrast 9: SStDSp < DStDSp). BOLD amplitude decreases due to the repetition of sentences were only found in the left anterior STS and did not differ between the event‐related and blocked designs.

Figure 6.

Repetition effect of sentences (contrast 9: SStDSp < DStDSp) estimated from the block‐design data (blue) and the event‐related data sets (red). No supra‐thresholded voxels were found in the right hemisphere.

It is interesting to note that the cluster thresholding is hiding a fairly strong (but spatially smaller) BOLD amplitude change in a more posterior area of left STS, another in right anterior STS, and a slightly weaker priming effect in Broca's area. As a result, the GRF thresholded activation pattern of contrast 13 (Fig. 7) appears far more extensive than just sentence priming. This, however, is likely to be just a thresholding artefact, in that the extra statistical power due to voice priming is increasing average Z‐statistics within a magnitude and they are then spreading into each other, passing the spatial threshold, and appearing as novel priming effects, even though they really are not.

Figure 7.

Effect of the cognitive interaction: maximum effect of repetition suppression (contrast 13: SStSSp < DStDSp).

Interaction effects between sentence content and voice.

The simple cognitive interaction effect (i.e., overall effect of repetition priming) is shown in Figure 7 (contrast 13: SStSSp < DStDSp). Unlike the previous contrasts (e.g., contrast 4), there were more differences between the event‐related and blocked versions. Priming was more extensive in the event‐related design with significant reductions in BOLD signal seen in the left hemisphere along the length of the left (STS), in lateral HG, all subdivisions of ventrolateral prefrontal cortex (pars opercularis, pars triangularis, and pars orbitalis), and posterior intraparietal sulcus (IPS). In the right hemisphere, only lateral HG and anterior STS showed neural priming effects. On the other hand, in the blocked version the only significant priming was found in the left STS. Part of this apparent difference is an artefact of cluster thresholding, as the blocked version led to significant priming effects in pars opercularis in the left hemisphere and the anterior STS in the right hemisphere. Nonetheless, several regions showed priming effects in the event‐related, but not blocked, versions and these included lateral HG, pars triangularis, pars orbitalis, and the posterior IPS in the left hemisphere. Presumably these are related to the well‐known strategic processing differences associated with “blocking” priming trials [Mummery et al., 1999; Neely, 1991]. Significant negative interaction (contrast 12) was found in right Broca's area and left posterior STS/MTG. No significant activation is found when testing for positive interactions between the sentence effect and the speaker effect.

Improved inference using mixture modeling.

Commonly in FMRI, null‐hypothesis testing is used on a contrast image to label voxels, or clusters of voxels, as being “active” if they reject the null‐hypothesis at a given false‐positive rate (FPR). This depends on knowing the distribution for the relevant statistic (e.g., regression parameter, t‐statistic, or pseudo‐t) under the null hypothesis.

An alternative approach for inference is to use mixture modeling on the statistic of interest [Everitt and Bullmore, 1999]. This involves fitting a mixture of distributions to the histogram of the statistic of interest; in our case we use a Gaussian for the central nonactivation part of the data, and a gamma for the activation and possibly another for “deactivation” [Beckmann et al., 2003b; Woolrich et al., 2005]. By using mixture modeling, instead of asking the question “Is the activation zero or not?” we ask the question “Is the activation bigger than the overall background noise?” Adaptability in modeling the nonactivating (“null”) part of the distribution can also help to protect against violations of the modeling assumptions, such as poorly modeled noise structure. For example, the presence of large‐scale signal fluctuations that are not removed by filtering can result in a significant shift of the assumed background Gaussian distribution associated with nonactivated voxels. Furthermore, mixture modeling provides more inference flexibility compared with null hypothesis testing. One can either control FPR (or TPR, true positive rate) by using the “nonactivating” or “activating” distributions, respectively. Controlling the TPR may be of real importance when using FMRI for presurgical planning.

For the FIAC data, Figure 8 (blue) shows the result of thresholding the GLM spatial Z‐statistic image of contrast 6 (DStSSp < DStDSp) with a Gaussian/Gamma mixture model. Unlike GRF thresholding, which failed to show any activation, the mixture model approach identifies small areas in lateral HG bilaterally, consistent with these regions showing a preference for human voice processing [Belin et al., 2000], although a different voice‐repetition study identified a lateralized priming effect more anteriorly in the right STS [Belin and Zatorre, 2003].

Figure 8.

Estimated speaker effect when changing sentence content (contrast 6: DStSSp < DStDSp) from the event‐related data. No significant difference was found when null‐hypothesis testing with Gaussian field theory was used for inference. When using alternative hypothesis testing via mixture modeling (P > 0.5), however, some difference was found, shown here in blue. Even more significant voxels were reported when, in addition, the HRF modeling was enhanced through the use of constrained optimal linear basis sets (shown in red).

Improved Signal Modeling Using Basis Sets

The fact that contrast 6 (DStSSp < DStDSp), testing for the primary speaker effect of interest, did not show any significant activation clusters when tested against the null hypothesis of zero effect size potentially is influenced by a vast number of modeling inadequacies. In particular, the standard GLM methodology described above used a canonical model for the assumed hemodynamic response and does not allow for any variation in the shape of the response at different voxel locations (although small variations in the delay are taken into account by adding the temporal derivative).

In order to investigate the impact of the fixed hemodynamic response function (HRF) assumption we used FLOBS (FMRIB's Linear Optimal Basis Set), a tool based on the idea of generating optimal basis sets for use in HRF convolution in FMRI linear modeling [Woolrich et al., 2004b]. FLOBS allows the specification of sensible ranges for various HRF‐controlling parameters (delays and heights for the different parts of the HRF convolution kernel), generates lots of example HRFs where each timing/height parameter is randomly sampled from the range specified, and then uses principal components analysis (PCA) to generate a basis set that optimally “spans the space” of the generated samples. The resulting basis functions are then used in the GLM to obtain constrained HRF fits. This is achieved by reprojecting the basis set onto the original set of samples in order to “learn” priors on the expected means and covariances of the individual basis functions. Using these priors when fitting the OBS model to the data means that the basis set is prevented from creating implausible HRF shapes. Constraining the shape of the estimated HRFs means that the noise is less “randomly fit” by the OBS model, giving better separation between the null part of the final statistics map and the activation part, i.e., better activation modeling power. Final basis functions approximately correspond to the canonical HRF (first), its temporal derivative (second), and its dispersion derivative (third).

Accurate HRF modeling is much more important in event‐related analyses than with block designs. In order to test whether constrained basis sets could improve the FIAC analysis, we reran the event‐related analyses using three optimal basis functions. In order to estimate group mean effect using a summary of power (rather than shape) for each condition for each subject, we generated an F‐statistic (across the three basis functions) for each contrast of interest at first‐level, converted this into a pseudo‐Z, averaged across sessions, and then carried out OLS mixed effects analysis at the group level.

The constrained basis‐set fitting approach results in statistic images where the null distribution at the highest level is not known. Therefore, we carried out inference using mixture modeling, thresholding at the alternative hypothesis test threshold of P > 0.5. The inference was limited to F‐statistic values that also show a positive fit to the first basis function (corresponding to the canonical HRF). This is in order to reintroduce directionality information that is not represented in the F‐statistics. Final postthresholded activation patterns thus represent potential changes in size and shape. The results can be seen in Figure 8 (red). Compared to mixture‐modeling results obtained from the original GLM Z‐statistic images, the fit to a flexible HRF basis set model results in large areas of activation seen in the posterior, middle, and anterior regions of the STS, bilaterally.

While generally it is the case that “larger” activations do not imply improved sensitivity, it is important to note that the improvements presented above were obtained using techniques that carefully control the probability of false‐positive detection. Consequently, these results are validly comparable and not attributable to an increase in false‐positive rate. Additional activation is only found in plausible auditory association areas, unlikely to be due to a regional unspecific increase in false detections.

MODEL‐FREE DATA ANALYSIS

The GLM analysis approach described above explicitly tests for differences in the BOLD amplitude modulation under the different conditions. Differences in shape are typically not tested for, although the inclusion of the temporal derivatives (Fig. 2) or more flexible time series modeling (e.g., use of optimal basis sets) allow testing for significant differences in the hemodynamic delay and/or differences in the shape of the response between conditions. Although these types of analysis add to the flexibility of the GLM approach, model‐based analysis is always fundamentally limited by the restrictions of the model chosen.

Exploratory data analysis approaches such as PCA or independent component analysis decompose 2‐D data matrices (where 3‐D space is unwrapped into a single 1‐D vector of voxels, hence the original 4‐D FMRI data set is represented as a time × voxels 2‐D matrix) into a set of pairs of time courses and associated spatial maps. Each time course plus spatial map describes a different component in the data. An advantage of such approaches is the ability to detect unknown, yet structured spatiotemporal processes in neuroimaging data [Beckmann and Smith, 2004; McKeown et al., 1998] and the additional modeling flexibility potentially results in an increased sensitivity in the detection of signals of interest.

Tensor‐ICA for Analysis of Group FMRI Data

The application of standard independent component analysis (ICA) to FMRI data was originally limited to the analysis of a single session's dataset at a time. In order to generalize ICA to the investigation of multiple sessions, subjects, or subject groups, we extended the standard 2‐D ICA methodology to higher dimensions (tensor‐ICA; see Beckmann and Smith [2004] for a detailed description of the model and a comparison with other approaches to exploratory multisubject analysis). In the tensor‐ICA approach, the group FMRI data is simultaneously decomposed into a set of components that now characterize the spatial, temporal, and subject domains. Figure 9 shows a schematic illustration of this multidimensional decomposition technique. Note that spatial maps obtained from such a data‐driven decomposition are explicitly optimized to violate a null‐hypothesis of random regression coefficients and therefore cannot be tested for significant activation based on simple null‐hypothesis tests. Instead, the approach taken here assigns voxel‐wise probabilities using an alternative hypothesis testing approach based on Gaussian/Gamma mixture modeling [Beckmann and Smith, 2004].

Figure 9.

Schematic illustration of the model‐free multisubject tensor‐ICA approach: the data are represented as a time × space × subject block of data that are then decomposed as the outer product of matrices describing the different signal components in the spatial, temporal, and subject domains.

On activation data the tensor‐ICA approach has been demonstrated to extract plausible activation maps, time courses, and estimates of the signal variation across the population [Beckmann and Smith, 2005]. Furthermore, it provides a richer description of the data than model‐based techniques, showing additional processes of interest such as secondary activation patterns or artefacts.

Mean Effects Tensor‐ICA

For the FIAC data, this approach is primarily used to help investigate the temporal signal characteristics in the auditory system for the block‐design data. In particular, we are interested in further investigating the question of whether there is significant within‐block repetition suppression, i.e., if there is any evidence for a BOLD amplitude decrease within blocks under the different conditions. For the tensor‐ICA analysis, the preprocessed FMRI data for each subject were split into four different datasets, one for each stimulation condition, each containing 44 volumes. For each condition these were then combined across subjects before being fed into tensor‐ICA, resulting in a set of time courses, spatial maps, and subject/session modes. Each decomposition showed exactly one strongly stimulus‐related effect where the spatial map shows significant voxels in primary and secondary auditory cortices and the associated time courses show clear block structure of the four repeats. Figure 10 summarizes the results of the tensor decompositions. The spatial map shows the tensor‐ICA estimated map (red‐yellow) together with the original FEAT GLM map (blue) for the DStDSp condition (C4; overlapping areas are rendered in green). The boxplots show the estimated response sizes of the 30 sessions for each of the four different conditions. There appears to be weak evidence for an overall reduction of BOLD modulation under the SStSSp condition, although not statistically significant. The time course plot shows the “peri‐block response” as estimated over the four conditions and all four repeats (boxplot at every TR within the block show the distribution of the 16 estimated values around the median). While the estimated BOLD amplitude decreases within the block, there is no evidence of a change in this response curve in the different conditions, i.e., no evidence of a systematic difference in shape over and above differences in magnitude.

Figure 10.

Results from the tensor‐ICA decompositions of the block‐design data. For the spatial map DStDSp, red/yellow = tensor‐ICA, blue = GLM, green = both. The tensor‐ICA spatial maps were thresholded by transforming the spatial raw IC estimates into Z‐statistic maps (dividing raw IC estimates by the voxel‐wise residual standard error of the decomposition) and thresholding these values at a posterior probability level of “activation” (P > 0.5) based on a Gaussian/Gamma mixture model fitted to the distribution of spatial Z values.

Between‐condition analysis.

The specific questions of voice and/or sentence repetition suppression, which within the GLM framework are being investigated by means of differential contrasts, can be approached in two different ways. First, similar to the GLM, we can compare the estimated tensor‐ICA spatial maps directly by calculating contrast maps, i.e., by directly generating difference‐maps from the four primary activation maps, obtained as the primary activation component in each of the four tensor‐ICA decompositions. As an example, Figure 11A shows the contrast of the primary tensor‐ICA map calculated from the SStSSp condition and the primary map calculated from the data obtained during the DStDSp condition. Also shown is the GLM Z‐statistic image for contrast 13 (green).

Figure 11.

“Contrasts” of primary tensor‐ICA mean effects maps: (A) contrast 13 (SStSSp < DStDSp) calculated explicitly from the 2 primary tensor‐ICA maps (red‐yellow) together with contrast 13 from the GLM analysis (green); (B) “exploratory contrast” calculated by decomposing the 4 primary auditory maps from the initial tensor‐ICA decompositions using standard ICA. One of the resulting columns of the mixing matrix shows strong correlation with the standard GLM contrast 10, and the spatial map (C; red‐yellow) is shown on top of GLM mixed‐effects results for contrast 10 (C; green).

As an alternative to calculating the contrasts explicitly as difference images between primary tensor‐ICA maps, we can instead use an exploratory data analysis at the second level and run standard ICA on the four primary tensor‐ICA maps (concatenated into a 4 × N matrix) in order to identify maximally non‐Gaussian source processes and associated “exploratory contrasts.” Figure 11B shows one of the resulting images. The associated contrast vector (red line) is mostly similar to the contrast 10 vector (SStSSp + SStDSp < DStSSp + DStDSp).

Additional Tensor‐ICA Group Maps

Out of the 19 maps that the tensor‐ICA estimated from the multisubject data (the number of components was estimated from the Eigenspectrum of the data covariance matrix using the Laplace approximation to the model order for a probabilistic PCA model; Beckmann and Smith [2004]), a few spatial maps show nonstimulus‐related yet spatially structured BOLD signal fluctuations. These spatial patterns can also be identified in resting FMRI data [Beckmann et al., 2005] and various researchers have suggested that these signal variations are of neuronal origin and correspond to functional resting‐state networks (RSNs). Examples of these patterns from the decompositions of the FIAC data are shown in Figure 12A. Regardless of their underlying cause, these patterns are a major source of nonmodeled noise in FMRI. In particular, the visuospatial system (default‐mode network; Gusnard and Raichle [2001]; blue) shows strong negative temporal correlation with experimental stimulation. Figure 12B shows the estimated temporal characteristics corresponding to this estimated default mode network (inverted for display here); different grey‐scales show the response as estimated within the four different conditions. The solid black line shows the mean estimated response in the primary auditory system. In all four datasets the estimated temporal response of the “default‐mode” deactivation is strongly (negatively) correlated with the response of the primary auditory system (Pearson correlation r > 0.5).

Figure 12.

(A) Additional group maps, depicting structured BOLD fluctuations in functionally relevant cortical areas, such as medial visual cortex (yellow), lateral visual cortical areas (red), visual stream (green), motor (pink), and visuospatial attention (blue), often termed the default‐mode network [Gusnard and Raichle, 2001]. (B) Estimated temporal responses within the default‐mode network for the 4 different conditions (gray, shown inverted) together with the temporal response for the primary activation pattern (black; average over all 4 conditions).

DISCUSSION

This article has presented results obtained from applying various tools from FSL (FMRIB Software Library) to data from the repetition priming experiment used for the HBM'05 Functional Image Analysis Contest. We first discussed a standard GLM‐based analysis using the FEAT model‐based tool; this showed expected temporal lobe activation in the main conditions [Cardillo et al., 2004; Crinion et al., 2003], and also the reductions in activation upon repetition priming [Dehaene et al., 2001; Desimone, 1996; Vuilleumier et al., 2002]. We then showed the improvement in sensitivity for detecting priming effects (for example, with respect to the speaker effect) when using recent more advanced analysis tools for modeling and inference (constrained optimal HRF basis functions and alternative hypothesis testing via mixture modeling). Note that these improvements were obtained using techniques that control the probability of false‐positive detection, so that these results are validly comparable and not attributable to an increase in false‐positive rate. Note also that these improvements are regionally specific to relevant areas. Finally, we applied model‐free analysis to the FIAC data using tensor‐ICA, which generated various results not obtainable by model‐based analyses; we showed the spatial, temporal, and cross‐subject modes in the main experimental conditions, proposed two different approaches for investigating specific differential contrasts in such datasets, and showed several resting state networks found consistently across subjects.

One of the main goals of this competition was to comprehensively analyze a particular dataset in order to highlight how different analysis techniques can identify findings beyond those seen in a basic GLM analysis. Thus, it is of considerable interest that we do not replicate the asymmetries observed by Dehaene‐Lambertz and colleagues [2006] given the same data. That is, for repeated sentences by different speakers, we observed a reliable neural priming effect along the length of the STS, bilaterally. In addition, by modeling HRF variability and using mixture modeling for inference testing, we demonstrated a similar voice‐repetition neural priming effect that also ran the length of the STS, bilaterally. In other words, both sentence content and speaker's voice led to extensive reductions in auditory association areas in both hemispheres. Qualitatively, at least, there appears to be no difference in the response of the left and right STS to content vs. voices. There may, however, be a quantitative difference—at least for sentence content—as the repetition effects were larger on the left than right hemisphere. No such laterality difference was apparent for voice repetition. The effects of voice were smaller than those of sentence and therefore required more sensitive statistical techniques to identify. Nonetheless, these results call into question the notion that sentence content is primarily left lateralized, while speaker identity is right lateralized or even limited to more anterior regions of the STS [Belin et al., 2000; Dehaene‐Lambertz et al., 2006]. Instead, the results suggest that both content and voice modulate a single underlying sentence processing system that includes anterior, middle, and posterior parts of the STS and MTG, bilaterally.