Technical NoteEffect of initial fMRI data modeling on the connectivity reported between brain areas
Introduction
Functional neuroimaging has allowed to establish a functional segregation of the brain areas activated during a variety of tasks. More recently, functional ‘integration’ studies have described how functionally specialized areas, i.e., areas whose activity is modified during the realization of a particular task, interact within a large-scale neural network (Lee et al., 2003). These studies have used multivariate techniques which provide an analytical tool to understand integrated systems by the analysis of the covariance matrix computed from the activity of several brain areas. Principal component analysis (PCA) allows to decompose neuroimaging data into a set of modes (Worsley et al., 1997). The functional system associated with various factors can be identified by comparing the temporal expression of a few modes with the occurrence of behavioral events. PCA does not require the definition of an a priori model but does not provide insight about causal influence. One method to study these causal influences is structural equation modeling (SEM). SEM is a linear technique which tries to explain the observed covariance patterns under some (anatomical) model constraints (Büchel and Friston, 2000).
All these methods for analyzing the connectivity between brain areas are based on the computation of the correlation coefficients between the activities of different brain areas. The raw time series is either that of one voxel or represents a group of voxels (region of interest) for one or more subjects. A characteristic time series is sometimes defined by averaging or extracting the principal eigenvector by PCA (Bullmore et al., 2000, Kondo et al., 2004, Penny et al., 2004). The correlation coefficients are usually not computed from the raw time series, but rather from prepared time series often resulting from an initial modeling of the raw time series using the general linear model (GLM). In particular, the original time series have to be preprocessed to remove variations which have no interest: scanner gain or induced artefactual or physiological variations. The correlation coefficients are then computed from the adjusted time series, which include variations related to different sources: subjects, tasks, residual variation not explained by the GLM.… The sources of variations taken into account depend entirely on the model chosen in this initial processing step. Nevertheless, the pretreatments used to remove confound-induced variations exhibit a great variability in the literature and are often poorly described (authors often use ambiguous terms such as “adjusted values” without a clear description of the model used). Confound induced variability is generally removed after applying a GLM (Whalley et al., 2005) or filtering in the frequency domain (Lowe et al., 1998). Further adjustments are less consensual like block averaging (Zhuang et al., 2005) or standardizing of the time series (Kondo et al., 2004).
Some studies further introduce a distinction between time series computed from task-related (controlled) and task-independent (uncontrolled) variations. The rationale when using task-independent variations is to look at an ‘intrinsic’ connectivity, resembling the one observed in the resting state. This leads to another interpretation ambiguity because the method used to remove the task effect can be very different according the report: e.g. a GLM (Whalley et al., 2005) or ICA (Arfanakis et al., 2000). Another ambiguity arises in studies focusing on one condition of the experiment where the authors keep only a part of the scans as the time series considered can be different if the initial data processing is done before (Honey et al., 2002) or after (Just et al., 2004) extracting the part of data for one task.
More generally, selecting different variance partitions can have a profound effect, both qualitatively and quantitatively, on the sample covariances and the ensuing inferences about connectivity. In conventional analysis using the GLM, the model used and the variance partitions considered are generally rather well documented, whereas in connectivity studies, they are often not described explicitly. This technical note aims at highlighting the importance of specifying the particular covariance component that has been characterized in terms of functional or effective connectivity, and the importance of interpreting the results in relation to this specific partition. These points are studied with both simulated fMRI time series and a real fMRI study of biological motion.
Section snippets
Modeling of the original time series
A schematic general experimental design is represented in Fig. 1. Fluctuations of activity can result from effects at several levels (task-, subject-, session- or block-related variations) and are associated with interactions between different levels (Subject × Task, Task × Block...).
In an attempt to simplify the notation, we have restricted the design to an example including only the factors “task” and “block” as experimental and time effects factors, respectively. This does not hinder the
Study of the correlation matrix between the time series from different brain regions
Several time series can be used to compute the correlation matrix between the activities of different brain areas.
Fitted data: the time series are reconstructed by the product of the task-related part of the design matrix X and the corresponding beta values: it is equivalent to say that XCβC+ε is removed from the original time series. The resulting correlation matrix reflects the correlation for the task-induced variations. This correlation matrix can be calculated only if the task-related part
Analysis of the correlation matrix between simulated time series
We have generated 1000 time series mimicking an experiment including 120 scans, corresponding to 3 tasks, with 4 blocks per task, and 10 scans per block. For task 1 (task 3), the average signal value across all the scans concerned is 0.625 (1.625), but with different mean values for each of the four blocks, with values of 0.25, 0.5, 0.75 and 1 (1.25, 1.5, 1.75 and 2), respectively. For task 2, the average signal value is constant across blocks and equal to 0.5 (block values 0.5, 0.5, 0.5 and
Analysis of an fMRI experiment
For an example with real fMRI data, we used an fMRI experiment whose results have been previously reported (Chaminade and Fonlupt, 2003). In this experiment, subjects had to detect the movement of moving points representing a human shape or an object shape. For each subject, the time series included 378 scans in 3 sessions of 126 scans. The experiment included 7 tasks, with 9 blocks of each task and 6 scans in each block. Two regions of interest (ROIs), located in the dorsolateral prefrontal
Discussion
The first aim of imaging experiments is generally to explore differences of activity during different tasks. More precisely, the subjects perform n repetitions of each task, and the means of the n repetitions are compared. A subsequent question concerns the occurrence of a relation between the activities of two ROIs, and this question can be asked in two ways. First, the means of the variations can be compared and give a result which can be, for example, “activity of region X increases during
Conclusion
It is not possible to decide in a universal manner what are the right types of variations to consider in connectivity analysis because it is fully dependent on the cognitive question underlying the experiment. However, the chosen variance partition to be studied must be clearly identified to avoid misinterpretation of the cognitive answer.
Acknowledgment
The authors wish to thank Karl Friston for most valuable advice and comments during manuscript elaboration.
References (17)
- et al.
Combining independent component analysis and correlation analysis to probe interregional connectivity in fMRI task activation datasets
Magn. Reson. Imaging
(2000) - et al.
How good is good enough in path analysis of fMRI data?
NeuroImage
(2000) - et al.
Assessing interactions among neuronal systems using functional neuroimaging
Neural Netw.
(2000) - et al.
Effects of verbal working memory load on corticocortical connectivity modeled by path analysis of functional magnetic resonance imaging data
NeuroImage
(2002) - et al.
Cooperation of the anterior cingulate cortex and dorsolateral prefrontal cortex for attention shifting
NeuroImage
(2004) - et al.
A report of the functional connectivity workshop, Dusseldorf 2002
NeuroImage
(2003) - et al.
Functional connectivity in single and multislice echoplanar imaging using resting-state fluctuations
NeuroImage
(1998) - et al.
Modelling functional integration: a comparison of structural equation and dynamic causal models
NeuroImage
(2004)
Cited by (10)
Disrupted functional connectivity for controlled visual processing as a basis for impaired spatial working memory in schizophrenia
2011, NeuropsychologiaCitation Excerpt :Temporal autocorrelation associated with artifact variances in the BOLD time-series due to scanner gain and physiological or movement related artifacts were removed in the preprocessing described above (i.e., mean-based intensity normalization and ICA artifact removal procedure). BOLD time-series data was further preprocessed to remove another type of temporal autocorrelation associated with linear trends due to session-related signal variance using a GLM procedure (Caclin & Fonlupt, 2006). Temporal Pearson correlation coefficients between all pairs of the 13 ROI time-series in each task condition were computed as the index of interregional functional connectivity for each individual.
The developmental cognitive neuroscience of functional connectivity
2009, Brain and CognitionLow-dimensional embedding of fMRI datasets
2008, NeuroImageAssessing functional connectivity in the human brain by fMRI
2007, Magnetic Resonance ImagingCitation Excerpt :The common feature of most of the above techniques is a dependence on the data underlying the calculation of correlation or covariance matrices. The results of a connectivity analysis, in general, will depend on the sources of variance used [48,49]. The underlying variance may be a function of the deviation of a given measure across subjects, or the deviation of a repeated measurement within a subject across time.
Schizophrenia and frontal cortex: Where does it fail?
2007, Schizophrenia Research