Technical NoteMotion or activity: their role in intra- and inter-subject variation in fMRI
Introduction
In a legendary study by McGonigle et al. (2000), the size of the activated area passing a fixed statistical threshold was shown to vary dramatically over 33 examinations of the same subject. This, and other studies, has lead to the common interpretation that fMRI has a large inter-session variance, not only when sessions span over different subjects, but even when the different sessions are examinations of the same subject. While McGonigle et al. (2000) emphasised that inter-session variation should not be assessed on thresholded images as the result will be highly threshold dependent, they did not provide information on the statistical significance of the task-by-session interaction. The presence of a significant task-by-session interaction is problematic in several ways. In the field of non-clinical brain mapping, using second level group analysis, there will be a decrease of sensitivity due to increased unexplained variance. When fMRI is used for presurgical planning false-negative activations might lead to inadvertant surgical removal of normal tissue, resulting in greater disability. Conversely, false-positive activations could lead to incomplete resection of a tumour.
The potential contributors to the intra-subject variability are numerous (Genovese et al., 1997). Compared to the number of choices available during fMRI acquisition and preprocessing, only a limited number of studies have investigated the impact of acquisition parameters and pre-processing methods on the within-subject sessional variation in the fMRI signal.
Slice-orientation, as reported by Gustard et al. (2001), had a non-significant impact on the reproducibility of the fMRI signal when isotropic voxels and a simple motor paradigm were used. Spatial normalisation was shown to have a significant effect on the reproducibility of visual activation by some (Swallow et al., 2003), but not others (Miki et al., 2000). In two studies, the effect of including neighbourhood information on the intra-subject variability was investigated. Intra-subject variability was decreased when larger smoothing kernels were used (Rombouts et al., 1998) or when the four nearest in plane neighbours were included in the analysis (Yetkin et al., 1996). These results are in agreement with what has also been observed for inter-subject variability (White et al., 2001, Shaw et al., 2003). Unfortunately, reduction of variability by smoothing comes at the cost of reduced spatial resolution. Physiological noise correction (Hu et al., 1995) was found to increase the test–retest reproducibility at 4 T (Tegeler et al., 1999).
Motion artefacts are some of the most important contributors to fMRI signal, unrelated to neural activity (Hajnal et al., 1995). Much of their effects can be removed by realignment but residual movement artefacts that are not accounted for by standard rigid-body realignment still exist. A commonly used method (Rowe and Passingham, 2001, Salek-Haddadi et al., 2003) for correcting these residual movement artefacts is to include movement-parameters in the design matrix of a general linear model. If only the raw movement parameters (translations and rotations) are included, it is assumed that the effects are linear, and that movement in opposite directions result in opposite signal changes. This is not always the case. Consider for example the 1-dimensional case of a grey matter voxel lying between two white matter voxels. In this case, movement of the voxel in either direction will lead to a signal decrease. Using a Volterra expansion of the movement parameters, higher order and differential effects can also be modeled, including spin history effects (Friston et al., 1996). In the case of stimulus locked motion, the inclusion of movement parameters in the design matrix is likely to remove not only residual movement artefacts but true activation as well, since these effects are no longer uniquely associated with the paradigm regressor.
The purpose of the present study was to investigate the extent to which the inter-session variability of the task-related fMRI activation could be improved by modeling residual motion, by inclusion of a Volterra expansion of movement parameters in the general linear model. In contrast to previous studies which have used coincidence maps, multi-panel displays or other strongly threshold dependent measures, we formally assessed the significance of the task-by-session interaction using F-contrasts.
Section snippets
Experimental setup
Ten healthy volunteers (labelled A–J) were examined with three tasks of overt word generation: Categorical (generation of words from a specific category), Alphabetical (generation of words starting with a specific letter) and Semantic (generation of verbs associated with a specific noun). Two of the subjects (A and B) were examined 10 times. The alphabetical and categorical paradigms were presented in a boxcar design with active and baseline condition lasting 44 s each. In the semantic
Results
This study produced 18 separate analyses. We are therefore only able to show detailed results from a subset of these analyses, and the remainder will only be commented on.
Discussion
McGonigle et al. (2000) highlighted the problem of large intra-subject variability in fMRI. We have here shown that residual movement artefacts are indeed a large part of the problem. Additionally, we have presented a method for explicitly assessing the significance of inter-session variance by the use of an F-contrast.
In the present study, we modelled residual movement effects by including a Volterra expansion of motion-parameters in the design matrix of a GLM. Using this method, we were able
References (26)
- et al.
Classical and Bayesian inference in neuroimaging: theory
NeuroImage
(2002) - et al.
Mixed-effects and fMRI studies
NeuroImage
(2005) - et al.
Effect of slice orientation on reproducibility of fMRI motor activation at 3 Tesla
Magn. Reson. Imaging
(2001) - et al.
Variability in fMRI: an examination of intersession differences
NeuroImage
(2000) - et al.
On the use of caffeine as a contrast booster for BOLD fMRI studies
NeuroImage
(2002) - et al.
Within-subject reproducibility of visual activation patterns with functional magnetic resonance imaging using multislice echo planar imaging
Magn. Reson. Imaging
(1998) - et al.
Working memory for location and time: activity in prefrontal area 46 relates to selection rather than maintenance in memory
NeuroImage
(2001) - et al.
Evaluating subject specific preprocessing choices in multisubject fMRI data sets using data-driven performance metrics
NeuroImage
(2003) - et al.
Reliability of functional localization using fMRI
NeuroImage
(2003) - et al.
Anatomic and functional variability: the effects of filter size in group fMRI data analysis
NeuroImage
(2001)
Analysis of fMRI time-series revisited—again
NeuroImage
A general statistical analysis for fMRI data
NeuroImage
Nonlinear spatial normalization using basis functions
Hum. Brain Mapp.
Cited by (191)
Functional MRI evidence for reorganization of language networks after stroke
2022, Handbook of Clinical NeurologyCerebello-thalamic activity drives an abnormal motor network into dystonic tremor
2022, NeuroImage: ClinicalReward modulates the association between sensory noise and brain activity during perceptual decision-making
2020, NeuropsychologiaCitation Excerpt :These modulated regressors were orthogonalized using the Gram-Schmidt procedure with respect to the regressor preceding it (i.e. the Feedback(points) regressor was orthogonalized with respect to the unmodulated Feedback regressor and the Feedback(reward) regressor was orthogonalized with respect to the unmodulated regressor and the Feedback(reward) regressor). Since head-movement related artefacts pose a significant problem for the interpretability of MRI datasets, we included a Volterra expansion to the six rigid-body motion parameters as covariates of no interest in the GLM (see Lund et al., 2005). Thereby, in addition to the estimated motion parameters from the realignment step, this expansion includes linear and quadratic effects of the motion parameters and the linear and quadratic effects of the first derivative of the motion parameters, resulting in 24 covariates of no interest.
Evidence for Hierarchical Cognitive Control in the Human Cerebellum
2020, Current Biology