Elsevier

NeuroImage

Volume 19, Issue 4, August 2003, Pages 1303-1316
NeuroImage

Regular article
Region of interest based analysis of functional imaging data

https://doi.org/10.1016/S1053-8119(03)00188-5Get rights and content

Abstract

fMRI analysis techniques are presented that test functional hypotheses at the region of interest (ROI) level. An SPM-compatible Matlab toolbox has been developed that allows the creation of subject-specific ROI masks based on anatomical markers and the testing of functional hypotheses on the regional response using multivariate time-series analysis techniques. The combined application of subject-specific ROI definition and region-level functional analysis is shown to appropriately compensate for intersubject anatomical variability, offering finer localization and increased sensitivity to task-related effects than standard techniques based on whole-brain normalization and voxel or cluster-level functional analysis, while providing a more direct link between discrete brain region hypotheses and the statistical analyses used to test them.

Introduction

Most hypotheses that are addressed using functional magnetic resonance imaging (fMRI) are stated in terms of the specific functionality of brain regions of interest (ROIs). These regions are frequently defined based on their cytoarchitectonic structure (e.g., Brodmann areas) or anatomical landmarks such as sulci Rademacher et al 1993, Caviness et al 1996. It is widely acknowledged (though rarely measured) that there exists a considerable degree of intersubject variability in the shape and location of these regions. We begin this article by presenting evidence that standard normalization techniques only partially accommodate intersubject variability and that after a full-brain normalization procedure there exists a considerable degree of residual variability in the shape and location of regions defined based on anatomical markers.

Since most fMRI experiments are built on multiple-subject data, standard functional analysis techniques based on voxel-level statistics attempt to compensate for this variability by spatially smoothing the functional series after normalization. Smoothing not only attempts to compensate for divergent functional anatomy but also enforces the validity of standard statistical analysis based on Gaussian field theory (Friston et al., 1996). A troubling aspect of this solution is the loss of clear regional boundaries resulting from the smoothing of the BOLD response across neighboring but possibly functionally dissimilar regions. Furthermore, the sensitivity of the resulting statistical tests is expected to decrease with the extent of anatomical variability across subjects.

In the present work we take a different strategy and present a methodology for the analysis of functional data that focuses on the activation of specific brain regions of interest. The proposed methodology is based on the definition of subject-specific ROIs and the testing of functional hypotheses directly at the level of the (multivariate) whole region activation. In this way, we avoid the need for full-brain intersubject coregistration and, most importantly, the need for spatially smoothing the functional series. By providing confirmatory analyses at the level of ROIs, the proposed methodology serves as a more direct link between the initial research hypotheses stated in terms of the functionality of discrete brain regions and the functional analyses used to test these hypotheses. This confirmatory approach to regional functional analysis is expected to ultimately increase the replicability of fMRI experiments and facilitate a knowledge buildup from functional results.

The outline of the article is as follows. The following subsection introduces the motivation for the proposed ROI analysis methodology. Intersubject Anatomical Variability describes a tool for the definition of ROIs based on anatomical markers and illustrates the extent of intersubject anatomical variability in temporal lobe cortical regions on a set of nine subjects. Functional Analysis summarizes the proposed methodology for the functional analysis of regional imaging data. Simulations comparing the expected sensitivity of the proposed methodology to one based on intersubject full-brain normalization and voxel- or cluster-level analyses are presented under simulations. Finally, conclusion presents Monte Carlo simulations validating the proposed statistical functional analyses on a range of possible fMRI noise conditions.

A major issue in functional brain imaging is the identification of functionally equivalent regions. A starting conservative hypothesis is that across different subjects there are identifiable brain regions that subsume the same functionality. In order to identify these equivalent regions, cytoarchitectonic anatomy seems to provide a good starting point. In this way, recent imaging work has shown a correlation between micro- and macroanatomy. In a study comparing cytoarchitecture to topographically defined ROIs, Rademacher et al. (1993) found that topographical features provided reliable limits for the boundaries of primary cortical areas BA 17, 41, 3b, and 4. The study showed two classes of variability in the size and shape of architectonic fields: variability predicted by limiting topographical landmarks and variability not predicted by these landmarks. The prominence of the former led the group to conclude that mapping systems based on individual topographical markers are more reliable than template-based systems at framing cytoarchitectonic areas.

Tzourio-Mazoyer et al. (2000), in a survey of studies that focused on the relationship between cytoarchitecture and macroanatomy, found that the relationship described by Rademacher et al. holds for primary and some language-related cortices, but is less clear for higher level cortical areas. When looking for a correlation between microanatomy and function, however, the same authors concluded that functional boundaries are not determined solely by cytoarchitectonic anatomy. Rather, an array of microanatomical criteria combines to determine functional cortical fields. Given the ambiguous relationship between microanatomy and function and the lack of cytoarchitectonic information by anatomical imaging, we believe that mapping individual macroanatomical variations provides the best means by which to reliably compare function across subjects. Our goal is to establish structure–function relationships with respect to anatomically defined ROIs and to then use these results to identify “functional ROIs.” This approach is well within the capabilities of functional and anatomical imaging and has been demonstrated to be reliable Caviness et al 1996, Kennedy et al 1998. Finally, we believe this approach offers the best available means for elucidating the relationship between anatomy and function. In the present work we follow this approach of defining subject-specific ROIs based on anatomical markers in an attempt to identify functionally equivalent regions across subjects. Although it would be optimal to define these ROIs on the functional images, the amount of anatomical information resolvable in these images is judged to be too low for replicable ROI definition. We therefore use the anatomical images to define the ROIs at the cost of introducing a certain degree of variability when coregistering them to the functional images. Contrasting with this proposed anatomically based ROI definition, it would also be possible to define the subject specific ROIs based on functional results alone. In this case the functional data used for ROI definition should be independent of that data used to test the hypotheses of interest on the ROIs. Otherwise, the latter results would tend to capitalize on chance fluctuations in the statistics that are dependent on the voxel selection criteria. An a priori ROI definition (one that does not depend on the functional hypotheses being tested) precludes such “capitalization on chance” problems. The proposed ROI methodology presented in this article utilizes an anatomically based ROI definition, reflecting our confidence in the reliability of this methodology (see “Demonstration of anatomical variability” for a discussion on the interrater reliability of the parcellation process). However, the proposed ROI functional analysis techniques would also be directly applicable in the case of subject-specific functionally based ROI definitions.

Once these functionally equivalent regions have been identified, a second, but related issue in functional imaging is the spatial alignment of these regions. The question is whether there is a spatial mapping between two subjects’ functionally equivalent regions such that paired voxels are also functionally equivalent. Spatial normalization followed by across-subject voxel-level statistical analyses assumes that an approximation to such a mapping exists. To correct for variability in this voxel-to-voxel functional equivalence mapping, the functional data is typically spatially smoothed. In the present work we chose to depart from this voxel-level equivalence hypotheses for several reasons. First, for practical reasons, smoothing the functional data will tend to partially pool the responses from neighboring cytoarchitectonically distinct regions, confounding the interpretation of the results. Second, for theoretical reasons, there could be circumstances in which it is not conceptually possible to spatially map across different subjects the voxel paradigm-correlated responses. For example, ocular dominance “zebra” patterns in V1, although sharing several spatial properties across subjects, are not superimposable. The topological differences between these patterns suggest that it is unlikely that any “optimal” spatial normalization technique could make them superimposable. Across-subject normalization and/or spatially smoothing would, in these cases, cancel out the (differential ocular dominance response) effects. The functional analysis techniques for the proposed ROI methodology attempt to relax this voxel-equivalence hypothesis by (a) providing a means of defining the intersubject pooling strategy based on a signal-independent or signal-dependent spatial basis and (b) providing multivariate statistical analyses that test the presence of an effect across multiple components of this spatial basis. By limiting the analysis results to voxels within a given ROI, the interpretation of these results is more straightforward, and by providing statistical tests on the regional multivariate response, an initial step toward a generalization of the voxel-level equivalence hypotheses is attempted. Conceptually, the ROI methodology attempts to make the basic units of functional analysis not individual voxels, where functional equivalence is most arguable, but the ROIs.

Section snippets

Definition of subject-specific regions of interest

A Matlab-based interactive toolbox was developed for the parcellation of ROIs from structural MRI scans. This toolbox currently allows for semiautomated ROI identification based on anatomical markers and is available for free download at http://cns.bu.edu/∼speech/ along with the functional analysis tools described in the next section. Although the exact relationship between anatomical markers and brain functionality is still open to discussion, anatomical markers provide a reliable and

Functional analysis

The relatively large extent of anatomical variability demonstrated above raises concerns about the statistical power and replicability of functional analyses based on standard whole-brain normalization. We believe a more robust approach for testing functional hypotheses would be to define subject-specific ROIs expressing the focus of the researcher’s interest while accommodating intersubject variability in the expected loci of activation and then test the specific functional hypotheses on the

Simulations

To simulate the effect of the observed anatomical variability on standard functional analyses, Monte Carlo simulations were run for both a standard procedure involving intersubject coregistration followed by voxel- and cluster-level analyses and the proposed ROI methodology to analyze controlled simulated functional data for nine subjects. The subjects’ parcellated structural images were used as reference for the location of simulated fMRI data. Simulated functional runs consisted of a single

Validation

The proposed statistical analyses were validated using Monte Carlo simulations with varying noise sources. Unless otherwise stated, the simulated data consists of a single run with 128 scans (TR = 2 s) over an ROI of 8 × 8 × 8 voxels (voxel size = 3 mm). A standard set of noise model parameters was defined as follows. Low-frequency noise autocorrelation width was set to 25s (FWHM), and wide-band noise was added at each frequency with 1/7th the peak energy of the low-frequency noise (peak ratio

Conclusion

We have presented a novel methodology for the analysis of functional MRI combining subject-specific ROI definition with a region-level multivariate statistical analysis technique. The proposed ROI-level analyses avoid the need to perform intersubject full-brain coregistration and spatial smoothing of the functional series. The methodology allows testing of functional hypotheses regarding the overall activation of specific regions of interest and hypotheses regarding the spatial profile of

Acknowledgements

We thank Dr. Julie Goodman, George Papadimitriou, Dr. David Kennedy, and Dr. Mukund Balasubramanian for their help and the MGH NMR Center for the use of their facilities. The study was supported by NIH Grants R29 DC02852 and R01 DC02852 (Frank Guenther, PI).

References (18)

There are more references available in the full text version of this article.

Cited by (138)

  • Broca's Area Is Not a Natural Kind

    2020, Trends in Cognitive Sciences
View all citing articles on Scopus
View full text