Spatial ICA reveals functional activity hidden from traditional fMRI GLM-based analyses

Independent component analysis (ICA) isasignalprocessingtechniqueusinghigher-orderstatisticstoextractsignalsbyunmix-ingsignalmixtures.McKeownetal.(1998)introduced spatial ICA (sICA) into func-tionalmagneticresonanceimaging(fMRI)study in the late 1990s. SICA assumes thatfMRI signal from each voxel represents alinear mixture of source signals, separatesthis signal mixture into spatially inde-pendent source signals using higher-orderstatistics, and groups all brain regionsshowing synchronized source signals intoindependent components (ICs), whichrepresent temporally coherent functionalnetworks(FNs)(McKeownandSejnowski,1998;McKeownetal.,1998;Calhounetal.,2002, 2009). McKeown et al. (1998) pre-dicted that sICA would be more sensitivein detecting task-related changes in fMRIsignal than the traditional general lin-ear model (GLM) based analysis, becausesICAusesadata-drivenapproach,andcanreduce noise in the ﬁnal solution by sepa-ratingartifacts from real fMRI signal.SICA hasbecome widely used for fMRIanalysis since its original application tofMRI 15 year ago (Calhoun and Adali,2012). Most studies use sICA as a tech-nique for extracting FNs from fMRI data,whilesomestudiesusesICAtoseparateand remove artifacts from real fMRI sig-nal for improving the sensitivity of sub-sequent GLM based analysis (e.g., Aronand Poldrack, 2006). Several studies com-pared sICA and GLM based analyses andreported that sICA revealed more brainregions showing task-related activationrelativetoGLMbasedanalysis,supportingthepredictionofMcKeownetal.(Calhounet al., 2001; Malinen et al., 2007; Tie et al.,2008; Kim et al., 2011). Relevant to thisissue, an interesting ﬁnding from sICAis overlap of two or more FNs of differ-ent timecourses and different task-relatedmodulations. This ﬁnding has been inter-preted as evidence of multiple concurrentcentral processes associated with commonbrain regions (Calhoun et al., 2008; Kimet al., 2009a,b; Menz et al., 2009; vanWageningenetal.,2009; Wuetal.,2009; StJacquesetal.,2011;Domagaliketal.,2012;Zhang and Li, 2012). Two recent studies,onefromus(Xuetal.,2013) andtheotherfrom Beldzik et al. (2013) systematicallyanalyzed FN overlap, and the task-relatedmodulation of the timecourses of over-lapping FNs. These analyses further high-lighted the presence of brain functionalactivity hidden from the traditional GLMbased analysis.In Xu et al. (2013),sICAwasusedto extract FNs from fMRI data relatedto a visual target-identifying task. Most(∼78%) functional brain regions showedoverlap of two or more FNs. Several brainregions including the dorsal anterior cin-gulate (dACC), insula and adjacent ven-trolateral prefrontal cortex (PFC), lateraltemporal and parietal cortex, and pre-cuneus/posterior cingulate (PCC) showedoverlap of seven or more FNs. Each ofoverlapping FNs showed unique task-related modulations of timecourses, andsomeofthemwere opposite(i.e.,increasesvs.decreases)toeachother.Thepredictionthat such opposite modulations withinthesamevoxelsmightbehiddenfromaGLM based analysis due to cancellationwas tested by analyzing the same fMRIdata using a GLM based analysis. Thebrain regions showing task-related activa-tion and deactivation as revealed by theGLM based analysis are 11.9 and 26.2%,respectively, of brain regions showing cor-responding changes as revealed by sICA,therefore supporting the prediction.In Beldzik et al. (2013),theauthorsused both sICA and a GLM-based analysistoassessanfMRIdatasetrelatedtoananti-saccadic task. They developed a tool calledContributive Sources Analysis (CSA) forestimating the amplitude of fMRI sig-nal changes in each FN. They ﬁrst usedthe GLM-based analysis to deﬁne clustersshowing signiﬁcant task-related activity asregions of interests (ROIs) and then usedCSA to extract measures of task-relatedchanges in fMRI signal within these ROIsfrom all FNs overlapping with these ROIs.They conﬁrmed their prediction that thesum ofthese measures from all FNsequalsto the measure of task-related changes offMRI signal within the ROIs as assessedby the GLM-based analysis. They fur-ther demonstrated that task-related oppo-site modulations of overlapping FNs con-tributed to the negative ﬁndings of theGLMbasedanalysisatsome brainregions.Therefore, ﬁndings from both studies arecomplementary, and indicate that mul-tiple neural circuits, each with uniquetimecourse and task-related modulation,can occur concurrently within the same

Independent component analysis (ICA) is a signal processing technique using higherorder statistics to extract signals by unmixing signal mixtures.  introduced spatial ICA (sICA) into functional magnetic resonance imaging (fMRI) study in the late 1990s. SICA assumes that fMRI signal from each voxel represents a linear mixture of source signals, separates this signal mixture into spatially independent source signals using higher-order statistics, and groups all brain regions showing synchronized source signals into independent components (ICs), which represent temporally coherent functional networks (FNs) (McKeown and Sejnowski, 1998;Calhoun et al., 2002Calhoun et al., , 2009.  predicted that sICA would be more sensitive in detecting task-related changes in fMRI signal than the traditional general linear model (GLM) based analysis, because sICA uses a data-driven approach, and can reduce noise in the final solution by separating artifacts from real fMRI signal.
SICA has become widely used for fMRI analysis since its original application to fMRI 15 year ago (Calhoun and Adali, 2012). Most studies use sICA as a technique for extracting FNs from fMRI data, while some studies use sICA to separate and remove artifacts from real fMRI signal for improving the sensitivity of subsequent GLM based analysis (e.g., Aron and Poldrack, 2006). Several studies compared sICA and GLM based analyses and reported that sICA revealed more brain regions showing task-related activation relative to GLM based analysis, supporting the prediction of McKeown et al. (Calhoun et al., 2001;Malinen et al., 2007;Tie et al., 2008;Kim et al., 2011). Relevant to this issue, an interesting finding from sICA is overlap of two or more FNs of different timecourses and different task-related modulations. This finding has been interpreted as evidence of multiple concurrent central processes associated with common brain regions (Calhoun et al., 2008;Kim et al., 2009a,b;Menz et al., 2009;van Wageningen et al., 2009;Wu et al., 2009;St Jacques et al., 2011;Domagalik et al., 2012;Zhang and Li, 2012). Two recent studies, one from us (Xu et al., 2013) and the other from Beldzik et al. (2013) systematically analyzed FN overlap, and the task-related modulation of the timecourses of overlapping FNs. These analyses further highlighted the presence of brain functional activity hidden from the traditional GLM based analysis.
In Xu et al. (2013), sICA was used to extract FNs from fMRI data related to a visual target-identifying task. Most (∼78%) functional brain regions showed overlap of two or more FNs. Several brain regions including the dorsal anterior cingulate (dACC), insula and adjacent ventrolateral prefrontal cortex (PFC), lateral temporal and parietal cortex, and precuneus/posterior cingulate (PCC) showed overlap of seven or more FNs. Each of overlapping FNs showed unique taskrelated modulations of timecourses, and some of them were opposite (i.e., increases vs. decreases) to each other. The prediction that such opposite modulations within the same voxels might be hidden from a GLM based analysis due to cancellation was tested by analyzing the same fMRI data using a GLM based analysis. The brain regions showing task-related activation and deactivation as revealed by the GLM based analysis are 11.9 and 26.2%, respectively, of brain regions showing corresponding changes as revealed by sICA, therefore supporting the prediction.
In Beldzik et al. (2013), the authors used both sICA and a GLM-based analysis to assess an fMRI dataset related to an antisaccadic task. They developed a tool called Contributive Sources Analysis (CSA) for estimating the amplitude of fMRI signal changes in each FN. They first used the GLM-based analysis to define clusters showing significant task-related activity as regions of interests (ROIs) and then used CSA to extract measures of task-related changes in fMRI signal within these ROIs from all FNs overlapping with these ROIs. They confirmed their prediction that the sum of these measures from all FNs equals to the measure of task-related changes of fMRI signal within the ROIs as assessed by the GLM-based analysis. They further demonstrated that task-related opposite modulations of overlapping FNs contributed to the negative findings of the GLM based analysis at some brain regions. Therefore, findings from both studies are complementary, and indicate that multiple neural circuits, each with unique timecourse and task-related modulation, can occur concurrently within the same voxels/regions, and that this feature of brain functional organization may not be detected by GLM based analysis. This observation is consistent with several lines of evidence in the literature.
First, anatomical studies indicate that each voxel may contain multiple signal sources of different timecourses. For example, the human cortex has about 25-30,000 neurons/mm 3 (Roth and Dicke, 2012), and each voxel may contain more than hundred thousands of excitatory and inhibitory neurons (Druga, 2009). They form multiple overlapping microcircuits. For example, each barrel column of the rat somatosensory cortex is about 300-500 µm in diameter and contains about 10 microcircuits (Brecht, 2007;Lubke and Feldmeyer, 2007). It receives two thalamic inputs of different response latencies, and generates five to seven outputs of various dynamics to different targets (Brecht, 2007). Therefore, it is reasonable to expect different microcircuits show different timecourses during processing of sensorimotor inputs. We acknowledge that whether and how the different timecourses of different microcircuits lead to different hemodynamic responses is not clear and should be investigated in future studies.
Second, electrophysiological studies indicate that neurons of different timecourses intermixed with each other in the same brain regions. For example, during the delay period of a working memory task, neurons in the dorsolateral PFC of monkeys show several patterns of activity changes including sustained increases in activity, increased activity at the beginning and gradually reduced activity later, and gradually increased activity (Chafee and Goldman-Rakic, 1998;Takeda and Funahashi, 2007;Verduzco-Flores et al., 2009). Furthermore, other neurons show various patterns of reduced activity during the delay period. It has been suggested that neurons showing different patterns of activity changes are associated with different cognitive processes such as cue representation, memory maintenance, and response preparation (Chafee and Goldman-Rakic, 1998;Verduzco-Flores et al., 2009). Based on these findings, Fuster hypothesized that different FNs overlap with each other in the brain (Fuster, 2009).
Third, published fMRI studies using analyses other than sICA report FN overlap. Yan et al. used a connected iterative scan (CIS) approach, which assesses FN overlap by detecting the involvement of the same functional clusters in different FNs, to assess FN overlap from fMRI acquired at resting condition (Yan et al., 2011). They reported overlap of the DMN with the so-called task-positive network at the PCC and lateral parietal cortices, consistent with the finding of overlapping FNs showing opposite modulation by sICA (Beldzik et al., 2013;Xu et al., 2013). A recent study applied temporal ICA to fMRI data and found that the DMN consists of multiple overlapping sub-networks (Smith et al., 2012). Several studies used seedbased approach to assess FNs from fMRI data and reported FN overlaps (Fox et al., 2005;Buckner et al., 2009;Tomasi and Volkow, 2011). The FN overlap reported in these studies may reflect the phenomenon of the timecourse of fMRI signal mixture in one region significantly correlating with the timecourses of two or more other regions, while the timecourses of the other regions do not significantly correlate with each other. However, the overlapping sICA components reflect different source signals from one brain region correlating with different source signals from two or more other brain regions.
The knowledge of FN overlap and their task-related modulation has important theoretical implications in understanding brain functional organization, and can help reconcile inconsistent data existing in the literature. For example, fMRI studies often show task-induced deactivation in the medial PFC and precuneus/PCC during tasks demanding working memory or cognitive control. It has been hypothesized that these brain regions are the core of DMN and associated with intrinsically generated, task-independent mental activity, e.g., task-unrelated thoughts (Fox et al., 2005;Raichle and Snyder, 2007). However, other findings indicate that these brain regions are also related to working memory. It has been reported that lesions of the medial PFC impair performance on working memory tasks (Tsuchida and Fellows, 2009;Barbey et al., 2011;Szatkowska et al., 2011). Furthermore, performance on working memory task correlates with the strength of functional connectivity within the DMN, especially between the medial PFC and precuneus/PCC (Hampson et al., 2006). The findings from the two sICA studies (Beldzik et al., 2013;Xu et al., 2013) of overlap of FNs with opposite task-related modulation indicate that there is no sharp border between neural substrates associated with cognitive control vs. task-unrelated thoughts, and that these neural substrates intermix with each other at both lateral and medial PFCs and precuneus/PCC. However, the lateral PFC may have more neural substrates associated with cognitive control, and fewer neural substrates associated with task-unrelated thoughts, and therefore, shows a net increase in fMRI signal mixture during task performance. On the other hand, the medial PFC may have more neural substrates associated with task-unrelated thoughts and fewer neural substrates associated with working memory, and therefore, shows a net decrease in fMRI signal mixture during task performance. This notion is supported by the electrophysiological evidence that neurons of one modality tend to concentrate in certain regions, but also distribute in other regions dominated by neurons of other modalities (Fuster, 2009).
In addition to the theoretical significance, the knowledge of FN overlap has important practical implications in how to assess brain functional organization. For example, seed-based approach is another popular technique for assessing FNs. This approach generates FNs by extracting the timecourse of fMRI signal from a selected voxel or ROI and then correlating this timecourse with the timecourses of other voxels or ROIs. Therefore, it does not consider different timecourses of different source signals from the same voxel or ROI, and thus may generate findings different from those by sICA, because some source signals from different ROIs may show significant correlations while the signal mixtures from these ROIs may show no significant correlations. Another example is that fMRI studies using GLM based analysis often report different changes in fMRI signal at some brain regions during task performance between patients and healthy controls. This difference in signal mixture Frontiers in Neuroscience | Brain Imaging Methods August 2013 | Volume 7 | Article 154 | 2 could be due to changes in one or more source signals. The knowledge of changes in a specific source signal may provide further insight into etiology and neuropathology of the disease under investigation relative to the knowledge of changed signal mixture.
In summary, sICA can separate source signals from the same brain regions, and this specific capacity has not been fully exploited in most published fMRI studies. Two recent studies explicitly employed this capacity and revealed new insight into functional activity hidden from a GLM based analysis. These new insights have significant theoretical and practical implications in understanding brain functional organization. Thus, we recommend that investigators should use sICA or some other approach which can account for overlap cancelation to perform secondary analysis of published fMRI studies analyzed by GLM alone, and use both GLM and sICA in future fMRI studies for a more complete understanding brain functional organization.

ACKNOWLEDGMENTS
This study was funded by the following grants: National Institute On Drug Abuse (NIDA) grants K01 DA027750 (Jiansong Xu) and P20 DA027844 (Marc N. Potenza).