Classification of fMRI independent components using IC-fingerprints and support vector machine classifiers
Introduction
Non-inferential or exploratory multivariate methods are being increasingly used in fMRI data analysis. These methods provide a characterization of the data, which does not rely on the statistical testing of a few stringent hypotheses and generate potentially valuable information on the nature of signal and noise in the fMRI time series. The value of such information consists in being complementary to that of statistical inferential maps. In some cases, however, the amount of information generated by these exploratory methods may be overwhelming and not easily interpretable.
In spatial Independent Component Analysis (ICA), for example, fMRI time series are decomposed into a large number (up to the number of scans) of spatial modes (independent components, [ICs]), with associated time courses (McKeown et al., 1998). In most cases, some of these components reflect interesting spatiotemporal patterns of stimulus-induced or spontaneous brain activity; other components reflect signal artifacts or noise (McKeown et al., 2003). The basic assumption in spatial ICA is that fMRI time series can be modeled as linear mixtures of latent sources, which can be blindly recovered under the constraint that their spatial distributions are mutually statistically independent. Several recent methodological and applied contributions indicate that this approach outperforms Principal Component Analysis (PCA) and can be a useful complement to standard hypothesis-driven univariate analysis (McKeown et al., 1998). In contrast to PCA, however, in which extracted components are naturally ordered according to explained variance, ICA does not provide any intrinsic order of the ICs. The experimenter is thus confronted with the problem of selecting and interpreting a subset of ‘interesting’ and ‘meaningful’ components.
In previous fMRI applications of ICA, selection of interesting components has been performed using various approaches. The simplest approach relies on the visual inspection of IC spatial maps/time courses (Bartels and Zeki, 2005, Calhoun et al., 2001a, Calhoun et al., 2001b). Selection of ICs based on their visual inspection, however, is very time consuming and highly dependent on the experience of the researcher. In most cases, ICs have been selected according to the amount of linear correlation of their time course with a model of the expected responses (McKeown et al., 1998, Schmithorst and Brown, 2004, Moritz et al., 2005) or related measures in the temporal frequency domain (Moritz et al., 2003). These approaches, however, appear to contrast with the data-driven nature of ICA. As an explorative tool, ICA may be particularly useful for detecting patterns of activity whose temporal dynamics cannot be easily modeled, such as in the case of hallucinations (van de Ven et al., 2005), epileptic seizures or in sensory or cognitive paradigms in which expected hemodynamic responses may be very diverse (Duann et al., 2002, Formisano et al., 2004, Castelo-Branco et al., 2002). Furthermore, ICA is being increasingly used for the study of ‘resting state’ functional connectivity (van de Ven et al., 2004, Greicius et al., 2003, Greicius et al., 2004) or as a de-noising step, which requires the selection of components reflecting noise and artifacts (Thomas et al., 2002). In all these cases, selection of ICs based on strong expectations on the profile of the IC time courses is insufficient.
Alternatively, selection of ICs has been performed using strong a priori assumptions on the spatial layout of the activation (Castelo-Branco et al., 2002, van de Ven et al., 2004). In this approach, distributed brain networks are detected by selecting ICA components that load heavily in predefined regions of interest (ROIs). A priori expectation on one or more ROIs, however, is not always available and, as in ROI-based univariate analysis, interesting processes occurring outside the predefined ROIs are ignored.
Other post hoc measures obtained from estimated ICs have been used for their sorting/selection. In analogy to PCA, McKeown et al. (1998) sorted the ICs according to their variance contribution to the original mixture. In fMRI data, however, neurophysiologically interesting phenomena are usually weaker than some of the sources representing structured noise. Thus, ranking of the ICs in this way may be not informative. Formisano et al. (2002) characterized the ICs using a combination of three descriptive measures (kurtosis of the spatial distribution, one-lag autocorrelation of the IC time course and clustering of the IC's spatial layout). The underlying idea was that ‘meaningful’ components aggregate in clustered regions in the three-dimensional space defined by these three measures. This heuristic criterion proved to be effective in isolating task-related components in a simple paradigm without using stimulus timing information.
The aim of the present article is twofold. First, we introduce the IC-fingerprint, a visual tool that aids the experimenter in displaying and characterizing the ICs. An IC-fingerprint is a representation of the component in a multidimensional space of descriptive measures, which can be visualized as a polar diagram. In line with Formisano et al. (2002), the underlying assumption is that ICs reflecting similar process types (e.g., BOLD activation, structured noise, movement) have similar fingerprints. To preserve the data-driven nature of ICA and the generality of the approach, the descriptive measures that define the space of the fingerprints are post hoc estimates of global properties of the ICs and do not rely on strong temporal or spatial hypotheses.
Second, we formulate the problem of selecting ‘meaningful’ components in the more general context of their (automatic) classification. After transforming the ICs in the multidimensional space of fingerprints, this problem can be formulated as subdividing the ICs in maximally disjoint classes and finding the optimal separating set of boundaries (hypersurfaces). Many different (supervised and unsupervised) algorithms may be used for this purpose (see Mitchell, 1997). Here, we describe and validate a supervised method for the classification of the ICs based on least squares Support Vector Machines (ls-SVMs). SVMs refer to a class of machine learning algorithms introduced by Vapnik at the end of the 70s (Vapnik, 1979). Ls-SVMs are a variant of SVM which have been proved to be effective in many problems of classification and pattern recognition (Suykens et al., 2002).
We illustrate our approach in the context of a multisubject fMRI study with visual structure-from-motion (SFM) stimuli (Kriegeskorte et al., 2003). We show that the set of measures that defines the IC-fingerprints is informative with respect to the problem of selecting and classifying fMRI-ICs and allows a reliable detection of interesting activation patterns. Furthermore, we show that an ls-SVM classifier, which is trained with a small subset of data from one subject, can automatically classify these fMRI-ICs in all other subjects with high correspondence to an expert classification. Finally, we show that the same classification algorithm can be successfully applied, without re-training, to fMRI collected using magnetic field, acquisition parameters, stimulation modality (auditory versus visual) and timing (event-related versus block design) considerably different from the SFM experiment used for training.
Section snippets
General description of the approach
Fig. 1 illustrates schematically the proposed approach. Individual fMRI time series are decomposed using spatial ICA in sets of ICs (see ICA decomposition). Obtained ICs are ‘transformed’ in a multidimensional space using a number of descriptive parameters (eleven in the current implementation). These parameters are computed from spatial distributions, spatial maps and time courses of ICs and characterize the global statistical and spatiotemporal nature of the sources (see Characterization of
ICA analysis and characterization of the ICs
Fig. 3, Fig. 4 illustrate the results of the ICA decomposition in Subject 1 (run 1) in the visual SFM experiment. Together with IC maps, visualized on the flattened representation of the subject's cortex, and time courses these figures include the eleven-dimensional fingerprint of each IC.
As expected, a subset of components included ICs that were consistently related to the stimulation protocol (Fig. 3A). The spatial maps of these components encompassed a widespread set of visual ventral and
Discussion
We have illustrated a general approach for the characterization and classification of fMRI independent components. Differently from conventional univariate statistical analyses, in which a small set of predefined hypotheses is tested, spatial maps (and associated time courses) obtained in fMRI-ICA are determined solely by the intrinsic structure of the data. Such a data-driven analysis provides an attractive opportunity for a blind detection of potentially interesting spatiotemporal patterns
Conclusions
We illustrated and validated a technique for the automatic classification and the selection of relevant ICA components in fMRI data. Its most important feature is that it matches the hallmark of ICA, i.e., blind detection of unexpected, yet plausible and interesting, neural (BOLD) activation patterns. The proposed solution facilitates the use of ICA for the explorative analysis of complex fMRI data sets. In combination with an appropriate choice of specific measures and heuristics, a similar
Acknowledgments
The authors are grateful to Nikolaus Kriegeskorte and Bettina Sorger for kindly providing the experimental data and for insightful discussions, and to Jean-Baptist Poline and the MADIC's Team (Orsay, France) for making the FIAC data available. Financial support from NWO (MaGW-VIDI grant 452-04-330) to E.F. is gratefully acknowledged.
References (47)
- et al.
Brain dynamics during natural viewing conditions: a new guide for mapping connectivity in vivo
NeuroImage
(2005) - et al.
Novelty indices: identifiers of potentially interesting time-courses in functional MRI data
Magn. Reson. Imaging
(2000) - et al.
Tensorial extensions of independent component analysis for multisubject fMRI
NeuroImage
(2005) - et al.
Semi-blind ICA of fMRI: a method for utilizing hypothesis-derived time courses in a spatial ICA analysis
NeuroImage
(2005) - et al.
Real-time independent component analysis of fMRI time series
NeuroImage
(2003) - et al.
Independent component analysis of fMRI group studies by self-organizing clustering
NeuroImage
(2005) - et al.
Spatial independent component analysis of functional magnetic resonance imaging time series: characterization of the cortical components
Neurocomputing
(2002) - et al.
Cortex-based independent component analysis of fMRI time series
Magn. Reson. Imaging
(2004) - et al.
Removing the effects of task-related motion using independent-component analysis
NeuroImage
(2005) - et al.
fMRI evaluation of somatotopic representation in human primary motor cortex
NeuroImage
(2000)
Independent component analysis of functional MRI: what is signal and what is noise?
Curr. Opin. Neurobiol.
Independent component analysis applied to self-paced functional MR imaging paradigms
NeuroImage
Empirical validation of the triple-code model of numerical processing for complex math operations using functional MRI and group Independent Component Analysis of the mental addition and subtraction of fractions
NeuroImage
ICA of fMRI group study data
NeuroImage
Noise reduction in BOLD-based fMRI using component analysis
NeuroImage
Functional connections among cortical regions: topography of EEG coherence
Electroencephalogr. Clin. Neurophysiol.
The spatiotemporal pattern of auditory cortical responses during verbal hallucinations
NeuroImage
Reducing multiclass to binary: a unifying approach for margin classifiers
J. Mach. Learn. Res.
Spatial and temporal independent component analysis of functional MRI data containing a pair of task-related waveforms
Hum. Brain Mapp.
A method for making group inferences from functional MRI data using independent component analysis
Hum. Brain Mapp.
Activity patterns in human motion-sensitive areas depend on the interpretation of global motion
Proc. Natl. Acad. Sci. U. S. A.
Modern Spectral Analysis
An Introduction to Support Vector Machines and Other Kernel-Based Learning Methods
Cited by (226)
Associations between childhood chronic stress and dynamic functional connectivity in drug-naïve, first-episode adolescent MDD
2022, Journal of Affective DisordersAnomalous intrinsic connectivity within and between visual and auditory networks in major depressive disorder
2020, Progress in Neuro-Psychopharmacology and Biological PsychiatryProspective motion correction of fMRI: Improving the quality of resting state data affected by large head motion
2020, NeuroImageCitation Excerpt :We applied a visual classification of ICA results based on features of the IC-maps and time courses, such as the presence of large clusters, and the absence of abrupt temporal changes (Griffanti et al., 2017). The group-defined default mode, visual, auditory, both right and left central executive networks, and one artifact-related component induced by head movement were visually identified (Damoiseaux et al., 2006; De Luca et al., 2006; De Martino et al., 2007; Mantini et al., 2007) and applied as masks in the corresponding ICs of each participant, and the mean beta was extracted as a measure of network robustness (Sampaio-Baptista et al., 2015; Marins et al., 2019). The power spectra for the temporal course from the five Resting State Networks-related ICs mentioned above were calculated for all conditions (RSPMCOFF, RSPMCON, XLegsPMCOFF and XLEGSPMCON).