Research articleCortex-based independent component analysis of fMRI time series
Introduction
One of the major advantages of functional magnetic resonance imaging (fMRI) over the other brain mapping techniques is the potentiality to look at the relationships between brain anatomy and function noninvasively and on an individual basis. This unique feature of fMRI is fully exploited when the functional topography of the brain areas is analyzed in relation to an explicit anatomical representation of the subject's cortex. In such cases, maps of brain activation as obtained from the statistical analysis of the functional time series are visualized on a folded or morphed (inflated, flattened) computerized reconstruction of the subject's cortical surface (see, e.g., Refs. [1], [2], [3], [4], [5], [6]).
In this type of approach, anatomical and functional information is merged together only at the final stage of the data analysis stream and mainly for the purpose of visualization. Recently, individual anatomical constraints have been used at an earlier stage as an additional constraint for the statistical analysis of functional imaging data [7], [8], [9]. Goebel and Singer [7] used the reconstruction of the cortical surface to restrict the detection of functional activation only to those voxels of a functional data set that lie within the cortex. This subset of voxels represents about 20% of the more than 100,000 voxel time courses (TCs) that are usually recorded. When the cortex is the target of the investigation, this subset also contains all the relevant functional information. This simple approach has been shown to enhance the sensitivity and the specificity of conventional statistical methods by reducing the severity of the multiple-comparison problem in whole-brain fMRI experiments. Kiebel et al. [8] used more directly individual anatomical constraints in the analysis of fMRI time series by introducing the framework of anatomically informed basis functions (AIBF). AIBFs allow incorporating anatomical prior knowledge based on reconstructed gray matter surfaces into spatiotemporal models of the fMRI time series. Andrade et al. [9] showed that cortical surface-based smoothing of functional time series increases the sensitivity of conventional voxel-based approaches that rely on general linear model (GLM) parameter estimation.
In all these cases, methods employed for the functional analysis were univariate [7], [9] or multivariate [8] hypothesis-driven statistical methods that require an a priori specification of a temporal model of the effects of interest. Exploratory data-driven approaches that do not make assumptions about the time profile of the effects of interest offer a complementary perspective to the conventional analysis of fMRI time series [10]. This might be especially useful in those cases in which the event of interest is not predictable (hallucinations, epileptic seizures) and when the hemodynamic response is difficult to model, such as in event-related designs with complex cognitive tasks or in experiments with perceptually ambiguous stimuli (see below).
Among data-driven methods, independent component analysis (ICA, [11]) appears to be particularly promising for the analysis of fMRI data (see Ref. [12] for a recent review). In the first applications of ICA to fMRI data analysis [13], [14], the ICA variant used is spatial in that the observed 4D fMRI signals are modeled as linear “mixtures” of unknown spatially independent processes (e.g., BOLD signal changes related to the cognitive task, physiological pulsations, head movements, artifacts, etc.), each contributing to the data set with an unknown time profile. An adaptive ICA algorithm [15] was adopted to decompose the time series into spatial components (ICs), each having a unique TC. The decomposition process maximizes the spatial statistical independence of the components, the idea being that the new representation of the data (ICs/TCs) reflects the “unmixed” configuration of the original spatial processes. Less frequently, a temporal ICA variant has also been adopted [16], [17], [18].
In the spatial ICA (sICA), as proposed in McKeown et al. [14], the entire matrix of the fMRI time series is blindly decomposed. This matrix includes not only signals from the cerebral cortex, but also from other parts of the brain, including subcortical structures, white matter and ventricles. The resulting decomposition, thus, also models the dynamics of the signal in these other structures.
Here, we combine sICA with methods of cortex segmentation and reconstruction and restrict the sICA decomposition to the “cortical” subregion of the matrix. We use the mesh of the white matter/gray matter boundary, automatically reconstructed from T1-weighted MR images [19], to limit the spatial ICA only to those voxels of the T2*-weighted functional time series which are within a specified region with respect to the cortical sheet (cortex-based ICA, or cbICA). We expect this cortex-based approach to improve the separation and anatomical accuracy of the ICs that represent cortical cognitive activations for two reasons. First, the number of voxels included in the data matrix does not affect the maximal number of spatial components which can be obtained (it equals the number of time samples, i.e., functional scans). Exclusion of extracortical contributions to the signal data set, thus, allows using the same number of components, otherwise used to separate noninteresting processes (e.g., signal changes in the ventricles, near the eyes, imaging artifacts), only for the processes occurring on the cortical surface. Second, improvements are also expected because of consideration on how the observed spatial mixtures are influenced by the “nature” of the included signals. Signals from, for example, the ventricles or near the eyes are “uninformative” with respect to the signals on the cortex. Thus, their inclusion in the data matrix leads to an increase of the complexity of the mixtures (in terms of number of sources) but does not improve the estimation of the cortical sources. Conversely, the restricted (yet statistically acceptable) sample of spatial observations considered in the cortex-based approach is highly “informative” with respect to the “interesting” sources and may lead to a better estimation of their spatial distribution.
We illustrate the cbICA framework in the context of whole-brain fMRI block and event-related experimental designs and in an experiment with perceptually ambiguous stimuli, in which the subject's perceptual state is not known a priori. Functional MRI time series is decomposed blindly into cortical surface components (CSCs) that can be directly visualized on the folded or morphed representation of the cortex. Results are compared to the results of conventional methods of linear multiple regression and principal component analysis (PCA), as well as to anatomically unconstrained spatial ICA [14].
Section snippets
General description of the cbICA approach
The steps of the cbICA are schematically illustrated in Fig. 1. Input data sets consist of a high-spatial-resolution 3D anatomical volume and a functional time series of the same subject, which have been previously coregistered (see below).
The initial steps of the cbICA approach are the generation of a cortical mask of each subject and the selection of a restricted functional data set consisting of the TCs corresponding to the cortical voxels in the functional time series. First, a polygonal
Experiment 1 (blocked design)
Fig. 2A shows the components (referred to as CSC1 and CSC2) corresponding to the cortical response to the visual stimulation (objects presented in the right and left visual hemifield alternately). CSC1 and CSC2 are the two components whose TCs were most highly correlated (R=.76 and R=.73) with hemodynamic predictor functions computed on the basis of the stimulation protocol by a linear model. The component maps were projected onto the folded and flattened representation of the subject's cortex.
Discussion
Cortex-based ICA is an approach for the analysis of fMRI data that combines techniques for the reconstruction and morphing (inflation, flattening) of the cortex from anatomical MR images with sICA of functional time series. We have demonstrated the validity of this approach by analyzing various fMRI data sets collected with different experimental designs (blocked and event-related designs, ambiguous stimulation) and by comparing the results we obtained with those of conventional
References (36)
- et al.
Cortical surface-based analysis. I. Segmentation and surface reconstruction
Neuroimage
(1999) - et al.
Mapping visual cortex in monkeys and humans using surface-based atlases
Vision Res.
(2001) - et al.
Mirror-symmetric tonotopic maps in human primary auditory cortex
Neuron
(2003) - et al.
Anatomically informed basis functions
Neuroimage
(2000) - et al.
Plurality and resemblance in fMRI data analysis
Neuroimage
(1999) Independent component analysis, a new concept?
Signal Process.
(1994)- et al.
Independent component analysis of functional MRI: what is signal and what is noise?
Curr. Opin. Neurobiol.
(2003) - et al.
An efficient algorithm for topologically correct segmentation of the cortical sheet in anatomical MR volumes
Neuroimage
(2001) - et al.
Independent component analysis: algorithms and applications
IEEE Trans. Neural Netw.
(2000) - et al.
Sustained extrastriate cortical activation without visual awareness revealed by fMRI studies of hemianopic patients
Vision Res.
(2001)
Tracking the mind's image in the brain I: time-resolved fMRI during visuospatial mental imagery
Neuron
Tracking the mind's image in the brain II: transcranial magnetic stimulation reveals parietal asymmetry in visuospatial imagery
Neuron
Tracking cognitive processes with functional MRI mental chronometry
Curr. Opin. Neurobiol.
Latency (in)sensitive ICA. Group independent component analysis of fMRI data in the temporal frequency domain
Neuroimage
Spatial independent component analysis of functional magnetic resonance imaging time-series: characterization of the cortical components
Neurocomputing
Borders of multiple visual areas in humans revealed by functional magnetic resonance imaging
Science
The constructive nature of vision: direct evidence from functional magnetic resonance imaging studies of apparent motion and motion imagery
Eur. J. Neurosci.
Visualization and measurement of the cortical surface
J. Cogn. Neurosci.
Cited by (89)
Analysis strategies for high-resolution UHF-fMRI data
2018, NeuroImageMethods for cleaning the BOLD fMRI signal
2017, NeuroImageCitation Excerpt :These alternative approaches to isotropic spatial smoothing take full advantage of improvements in the spatial and temporal resolutions achievable with advanced sequences and acquisition protocols. For example, ICA-based denoising approaches benefit from a cortex-based analysis at the single-subject level (Formisano et al., 2004), particularly for high resolution fMRI studies (De Martino et al., 2011). For resting state fMRI data, it is best to perform temporal filtering (i.e. low-pass, high-pass or band-pass filtering), nuisance regression and censoring simultaneously in a unique regression model to minimize potential errors in data denoising.
Task-evoked brain functional magnetic susceptibility mapping by independent component analysis (χICA)
2016, Journal of Neuroscience MethodsLateralization and localization of epilepsy related hemodynamic foci using presurgical fMRI
2015, Clinical Neurophysiology