Frequently asked questions
From FMRI Methods Wiki
Here are a number of commonly asked questions with regard to the measurement and analysis of fMRI data. The current list of addressed issues is not exhaustive: please feel free to add or suggest additional questions, as well as to comment or provide extensions to suggested answers.
- What does fMRI measure?
The most common use of fMRI includes the measurement of the Blood Oxygen Level Dependent (BOLD) signal, a technique which was developed in the 1990s after the BOLD effect was first described (Ogawa et al., 1990). This effect reflects the dependence of T2* weighted contrasts on the amount of blood deoxygenation which is a consequence of different magnetic properties of oxygenated and deoxygenated hemoglobin. The measured BOLD signal is sensitive not only to the level of blood oxygenation, but also to closely related cerebral blood flow and cerebral blood volume.
Since oxygen consumption reflects metabolic activity which is dependent on the level of activation of the surrounding tissue, BOLD signal can be used as an indirect indicator of neural activity. BOLD response reflects primarily input and the local processing within a given brain area and not so much its output (it is more related to presynaptic than spiking activity; Logothetis et al, 2001). Logothetis et al. (2004) also showed the correlation between the density of vasculature and the number of synapses, not necessarily the number of neurons in a particular region.
- What are the main advantages and disadvantages of BOLD fMRI?
Main advantages of fMRI are its noninvasive nature and good spatial resolution. In comparison to other fMRI techniques, e.g., arterial spin labeling or the assessment of oxidative metabolic rate or blood volume, measurement of the BOLD signal is easier to implement and has higher functional contrast to noise expressed as the ratio of signal change and background fluctuations (Bandettini, 2006).
Its main disadvantage refers to the nature of the BOLD signal: it is only an indirect measure of neural activity which is also biased toward one type of neural processing, namely input and intraregional processing. In addition, it is susceptible to several imaging artefacts and has limited temporal resolution. Although this can be improved to a certain degree, increasing temporal resolution comes at the expense of decreasing the spatial one. Due to the rather low signal-to-noise ratio of fMRI, analysis of the acquired data is rather extensive, complex and occasionally based on rather simplified assumptions, e.g., those related to the properties of hemodynamic response.
- What artefacts can occur in the acquired fMRI data?
fMRI data is susceptible to a large number of artefacts which can roughly be divided into scanner-induced and physiological artefacts. Scanner-induced artefacts include e.g., B0 inhomogeneity, radiofrequency or gradient artefact. Physiological artefacts include e.g., motion or contamination from large veins and arteries in the brain. First of all, BOLD fMRI signal can be displaced a few millimeters from the area of neural activation and towards the large draining veins. This occurs because draining veins are not as oxygenated as large arteries and capillaries and are subject to greater decrease of deoxygenation than other blood vessels. In addition, since large veins drain rather extended areas of the brain, activations captured by hemodynamic methods are often more extended than the area of neural activation (spatial blurring). A big problem in fMRI includes signal dropout in some brain regions and partial volume artefact which depends on the voxel size.
Planning and conducting an fMRI study
- What is important to do before conducting an fMRI study?
Quite a few things need to be done before conducting an fMRI study. Most importantly, after specifying the research questions of interest, the research design has to be carefully planned and modified for the fMRI environment. Some features of this design as well as the technical equipment and desired scanning settings should ideally be tested before starting the experiment.
When specifying and checking the design of the study, it is very important to consider its efficiency (detection power and/or estimation efficiency). Generally, the power of the design and potential value of recorded data depend strongly on factors such as e.g., number of trials within each condition, separation of trial events (duration and variability of SOA), ortogonality / independence of parameters used within the design, randomization, etc. In addition, one should always conduct at least one behavioral study prior to the experiment to check if participants understand and are able to perform the task as well as to insure comparability between different experimental conditions (e.g., condition difficulty).
In addition, it is important to test technical equipment before starting data collection. This includes, for example, checking presentation and response recording software (important to relate scanner pulses to these), visibility and general perceptibility of stimuli. Checking how well the participants can perceive the presented stimuli is important because fMRI environment can sometimes be very different from the one in which the experiment was programmed or behaviorally tested.
- What are the main types of designs used in fMRI?
First of all, it is possible to distinguish designs which employ a certain (cognitive) paradigm consisting of one or more tasks performed by the subject and those which do not employ such a paradigm, e.g., studies of the resting state / default network / stimulus independent processing (although these can include a fixation task which is also sometimes considered as a proper task). Although most designs which include a predefined paradigm are fully controled by the experimenter, in some cases this may not be possible, e.g., if the occurence of certain events can only be indicated by the participant (e.g., perception of ambiguous objects) or when using natural stimuli paradigms (Hasson et al., 2004).
Main types of “paradigm designs” are block and event-related designs. Depending on the question of interest, it is possible to somewhat adjust these designs, as is done in e.g., parametric or rapid event-related fMRI designs. In addition, it is possible to use mixed block/event-related designs which allow the separation of transient and sustained activity in fMRI (Visscher et al., 2003).
- What are the main advantages and disadvantages of block designs?
The main advantage of block designs is their high statistical power: rapid repetition of stimuli of interests has an additive effect on the measured BOLD signal. The main disadvantage is their limited use, because block designs can only be exploited for some research questions due to their highly artificial nature, e.g., the fact that events within these designs are always predictable and non-randomized (for example, oddball paradigm can not be investigated using a block design). These designs are in principle limited to the use of the subtraction paradigm.
- What are the main advantages and disadvantages of event-related designs?
A disadvantage of event-related designs is their lower statistical power in comparison to block designs. Their main advantage is related to a much wider scope of research questions which can be addressed using this type of design (e.g., oddball paradigm, task switching or paradigms in which only the subject can indicate an event can be investigated exclusively using event-related designs). These designs allow inclusion and removal of certain types of trials, e.g., catch trials whose one potential purpose can be testing / ensuring participants' attention on the task at hand. Furthermore, they allow separation of trials based on participant's behavior, e.g., erroneous and correct responses within a certain condition. Most importantly, the fact that event-related designs allow randomization of events within an experiment is very important as it minimizes strategy, as well as anticipation and habituation effects.
- Is fMRI data always saved in the same format and, if not, what does this depend upon?
fMRI data can be saved in different formats (e.g., NIfTI, Analyze, vista, DICOM), depending on the scanner on which it is recorded as well as the software package used for its analysis. Luckily, it is possible to convert data from one format to another using different conversion algorithms which are available in most analysis softwares.
During conversion it is important to consider slice acquisition parameters of the conducted fMRI study as well as slice orientation of the recorded data. It is important to emphasize that the automatic conversion procedure can sometimes fail, usually because different formats employ different conventions of slice orientations. Specifically, most problems arise because some formats support radiological (left brain = right image) in contrast to others which employ the neurological convention (left brain = left brain).
Therefore, it is always important to check whether the conversion was performed correctly or not. While the anterior-posterior (y) and superior-inferior (z) axis can be checked rather easily, problems can arise with the left-right (x) axis. In order to correctly identify brain hemispheres, it may be useful to think in advance and somehow mark different sides of the body during scanning (e.g., it is possible to place a vitamin E pill to one side of the head). Alternatively, it is possible to introduce an additional short functional task which can later disambiguate the hemispheres (e.g., a short finger tapping task can be performed before starting the experiment) or use the information from response periods if participants had a task to which they responded using only one hand. Alternatively, some labs have access to additional anatomical scans of individual participants recorded in separate sessions which may also be used for this purpose.
Analysis of fMRI data
- What does the analysis of fMRI data include?
The analysis of fMRI data includes preprocessing and statistical analysis.
Preprocessing refers to a series of steps aimed at removing uninteresting variance (noise) which increases functional signal-to-noise ratio and preparing data for statistical analysis. These can roughly be divided into more anatomical and more functional steps. Anatomical ones include spatial coregistration and normalization, namely the transformation of individual participant fMRI data into standardized coordinate (MNI or Talairach) space. Functional ones include motion correction (spatial realignment), slice timing correction, spatial filtering (smoothing) and temporal filtering of the data.
Statistical analysis can include different procedures which can be differentiated primarily based on whether they include an underlying model (model-based vs. model free approaches) and whether the analysis is conducted independently for each voxel or not (univariate vs. multivariate analysis). Currently, the most commonly used approach is the model-based univariate analysis as implemented within the General Linear Model (GLM).
The choice of the statistical analysis of fMRI data depends on the research question motivating a study and can have repercussions on the choice of preprocessing steps / parameters conducted before performing the analysis itself.
- What softwares can be used for analyzing fMRI data?
Quite a few software packages are available for analyzing fMRI data. Some of them contain tools for all steps of standard fMRI analysis (spatial coregistration and normalization, preprocessing, statistical evaluation, visualization), while some specialize in only some aspects or special analysis approaches. Some are free and some commercial. Some are available for different operating systems and some limited to fewer platforms (e.g., only to Linux). Some are self-standing and some require other programs (e.g., Matlab). Different softwares also differ in types of algorithms they apply for different analysis steps, number of available features, speed of computations, simplicity of use, available support, etc. These and other factors should be explored and considered when choosing a preferred software. A non-exhaustive list of available softwares can be found here.
Spatial coregistration and normalization
- What is spatial coregistration?
Spatial coregistration refers to intrasubject registration, namely the alignment of functional and structural data from individual participants. This is usually done when reference anatomical data sets acquired in separate sessions are available for each participant. Spatial coregistration includes computing a transformational matrix specifying the transformation parameters and applying it to the data of interest.
- What is spatial normalization?
Spatial normalization refers to intersubject registration, namely the process of normalizing or scaling the data to a standard space. There are two widely accepted standard stereotaxic coordinate spaces, Talairach and MNI, which differ in the shape, size and representativeness of the provided templates. Spatial normalization can be done either by linear scaling in the x, y and z direction or nonlinear scaling which includes some additional, nonlinear deformations and is usually based on gross morphological landmarks. Normalization represents an important step of analysis as it allows group analysis and comparisons / generalizations of the obtained results.
- How comparable are the Talairach and MNI brain? What are the advantages of using one over the other?
Since the Talairach brain is based on one post-mortem brain of a 60-year old woman (with only one hemisphere, no brainstem or cerebellum), it has often been criticized for its suboptimal representativeness. As an alternative, Montreal Neurological Institute developed new standard template which approximately matched the Talairach brain and was more representative for the population. The first template was MNI305 after which ICBM152 and ICBM452 were introduced. Unlike these which are based on a large sample of participants, template Colin27 is based on 27 scans of one subject which was later aligned to the MNI305 template.
Although it is possible to convert Talairach to MNI space and vice versa, this is not a trivial problem because the two templates differ in shape and size: MNI brain is somewhat longer, higher and deeper and the biggest differences between the two occur further away from the middle of the brain. This conversion can be done using different, both linear and non-linear methods, most popular of which is the piecewise linear mni2tal procedure (developed by Matthew Brett).
Additional preprocessing steps
- How problematic are motion artefacts in fMRI?
Motion artefacts are very problematic in fMRI as they can seriously degrade the quality of the collected data. Next to the sub-optimal behavioral performance, they represent the most often reason for discarding data sets from the analysis. Although it would be ideal to prevent them from happening, this is not possible as it would include eliminating not only (partly) voluntary whole body / head movements, but also involuntary movement due to, e.g., breathing, blood flow or heart beats. However, it may still be possible to reduce this problem by giving detailed instructions to participants and using some non-radical immobilization strategies aimed at restraining such movement. Alternatively, motion artefacts can be corrected online during the acquisition of fMRI data with prospective motion correction algorithms or offline after the scanning. In the latter case, motion artefact correction or spatial realignment constitutes a standard step in the analysis of fMRI data and is aimed at realigning acquired images in a geometrically compatible fashion.
In most analysis softwares it is possible to get information regarding the extent of movement in single participants and / or the parameters which were used for correcting the data. These parameters can later also be accounted for in the statistical analysis if they are included as additional covariates of no interest. One can also usually visually inspect the collected fMRI time series and check if the head moved significantly during the course of the experiment. In cases of severe motion, it can happen that the realignment algorithms can not correct the artefact. If data sets with residual artefacts are further processed and statistically analyzed, the artefact can sometimes be visible in functional contrasts (for example, “aura-like” activations located at the rim of the brain could be an indicator of this problem).
- What is slice timing correction?
Every 3D volume of the entire brain consists of a number of slices which are collected at slightly different points in time. Therefore, time series from different voxels are not comparable unless this temporal offset is taken into account. One way to deal with this problem is to use slice timing correction algorithms which “temporally realign” all acquired slices to one reference slice (usually the first or the middle one) using interpolation. When performed, this preprocessing step is done either before or after spatial realignment.
The answer to the questions whether and when to use this correction depends on a number of factors, e.g., slice acquisition parameters (continuous vs. interleaved acquisition) or planned modeling of HRF which may or may not incorporate the use of a temporal derivative of the basis function.
- What type of filtering does the analysis of fMRI data include?
The analysis of fMRI data includes spatial and temporal filtering.
Temporal filtering is used for removing some frequencies which are considered as noise within the acquired signal. In addition to applying high-pass filter which removes low frequency drifts over time (mostly scanner of physiological noise), it is also possible to apply an additional low-pass filter in order to remove high frequencies which may sometimes be needed, e.g., in experiments with long trials or block designs. Specifying cut-off frequencies during temporal filtering is not always straightforward and it is important to be very careful not to remove significant portions of the signal-of-interest during this process.
Spatial filtering increases signal-to-noise ratio and reduces inter-subject variability. It is performed by applying spatial filters which usually use a Gaussian distribution. The size of the spatial filter is normally expressed as FWHM (full width at half maximum). The disadvantage of spatial filtering is a decrease in spatial resolution in the single data sets.
- What is needed for performing a statistical analysis of fMRI data?
Statistical analysis requires a set of preprocessed functional data and a design matrix which specifies the experimental design (specifying all conditions within the design and their respective onsets, durations and amplitudes reflecting potential parametric manipulations). Next, analysis parameters need to be chosen. In the standard GLM approach this would include choosing the basis set or function for modeling HRF as well as main analysis approach (e.g., standard GLM, Bayesian analysis) which can also include a decision regarding the manner autocorrelations within the fMRI time series should be handled (pre-coloring vs. whitening).
- What does General linear model (GLM) include?
GLM, the most commonly used type of analysis of fMRI data, aims at determining brain areas which are significantly activated in different conditions within the experiment. First of all, GLM includes modelling the BOLD response and specifying the parameters of the model. In addition, using GLM it is possible to determine whether an increase of the BOLD response occurred in response to a certain experimental manipulation (more specifically, it is possible to determine the probability that the measured change in BOLD could occur by chance, thereby allowing the estimation as to whether it significantly differs from zero or not). The thresholding of the statistic map obtained using GLM can be based on the point distribution of the statistic (uncorrected) or can be corrected for multiple comparisons using one of several available correction approaches.
- What is the hemodynamic response function (HRF) and how can it be modeled?
The hemodynamic response function (HRF) describes the way in which the BOLD signal evolves over time in response to a change in neural activity. A typical BOLD response comprises of a short-lasting initial dip, rise, peak (after 5 sec), fall and undershoot slightly below the baseline before returning to the initial value after 12-24 sec. HRF depends on the properties of stimuli and underlying neuronal activity and, although it is relatively stable across sessions recorded in same conditions, significant differences between individuals and different regions within individuals have been reported (Aquirre, Zarahn and D'Esposito, 1998).
There are several types of basis sets or functions which can be used for hemodynamic modeling, some of which may be more appropriate for certain types of experimental designs, e.g., box-car, half sine, discrete cosine, gamma, modified Gamma, Gaussian function, canonical HRF, Finite Impulse Response, Fourier basis sets. When modeling HRF, it is possible to use one or more basis functions for each voxel. Including additional basis function or adding partial derivatives can sometimes provide better models of the data as they can account for e.g., temporal offsets or dispersion of the data.
- Why is correcting for multiple comparisons important in fMRI and how is it performed?
Correcting for multiple comparisons is important because a standard fMRI analysis includes computing separate statistical tests for each voxel, which may amount to 100 000 tests or more, depending on voxel size. There are several ways in which such correction can be applied, e.g., using familywise error rate or false discovery rate procedures, cluster-size thresholding, nonparametric procedures or alternative analysis approaches which may not require such corrections.
Familywise error rate procedures control the chance of any false positive occurrence and include classical Bonferroni corrections or maximum distribution methods, e.g., Gaussian field theory and the permutation test. The classical and very conservative Bonferroni correction is based on the number of independent tests which equals to the number of measured voxels (calculated as a ratio of the probability threshold and the number of tests). Since this procedure can be quite stringent and lead to severe Type II (false negative) errors, using Gaussian Random field theory it is possible to recalculate and reduce the number of independent resells based on data smoothness. This is done under the assumption that the fMRI signal at each voxel has a normal spatial distribution.
False discovery rate is aimed at controlling the ratio of hits to false alarms and it adjusts the criterion used based on the amount of signal present in the data. One can also combine cluster-size thresholding with single voxel probability thresholding.
Alternatively, one can also reduce the number of comparisons by restricting either data acquisition to some parts of the brain or data analysis using region-of-interest statistical analysis. This is usually done when one can specify a strong a-priori hypothesis about the involvement of a certain brain area in a task of interest. In addition, some analysis approaches, e.g., multivariate or Bayesian analysis may not need such corrections.
- What are the benefits of using Bayesian analysis in fMRI?
Using Bayesian analysis in fMRI may provide a high efficiency alternative to classical modelling approaches. Practical benefits include its high tolerance against outliers and a non-equal influence of individual subjects on estimated group posteriors which is dependent on within-subject variance. In other words, the results of Bayesian analysis may be less prone to detrimental effects of “exceptional” or highly unstable (variable) data sets. In addition, it has been argued that correction for multiple comparisons is not necessary following Bayesian analysis which might also qualify as a potential benefit.
- What are model-free or exploratory fMRI analysis approaches?
Model-free or exploratory analysis are those which principally feature no or minimum assumptions about the hemodynamic response (they do not include an a priori specified model). In addition, they are primarily based on the internal structure of the data which does not have to be linked to the experimental design. Examples of model-free analysis include e.g., independent component analysis (ICA), similarity approaches (correlation), clustering techniques, etc.
- What is the difference between univariate and multivariate analysis of fMRI data?
The main difference refers to the way individual voxels are treated within these approaches. In univariate analysis each voxel is treated independently which is clearly an oversimplificated and quite reductionist assumption. This motivated the development of multivariate analysis techniques which take into account the activity of a larger number of voxels and can therefore have higher informational value in comparison to univariate techniques. There are numerous multivariate analysis techniques available (e.g., independent component analysis (ICA), partial least squares (PLS), clustering approaches) which can, based on the number of prior assumptions included in the analysis, roughly be divided into more exploratory and more confirmatory techniques.
Currently the most debated multivariate analysis approach is the multivariate pattern-based analysis (also known as multivoxel pattern recognition, brain reading, multivariate information-based mapping, pattern-information fMRI or MVPA: multi-voxel pattern analysis) which includes an application of complex pattern-classifier algorithms to multivoxel patterns of activity with the goal of decoding its representational content. There are several different decoding approaches available, among which linear decoding is most often used.
- What are the main connectivity analysis approaches currently available?
In addition to studying anatomical connectivity, interregional interactions can be investigated using functional imaging by measuring functional and effective connectivity between different brain regions. Functional connectivity refers to identifying covariance between different regions and can be investigated using e.g., interregional correlations, psychophysiological interactions (PPI), principal and independent component analysis (PCA and ICA) or multivariate autoregressive (MAR) models. In contrast to functional connectivity which does not imply causality, effective connectivity goes beyond pure description and tries to identify the influence that one system exerts on another. Effective connectivity can be assessed using structural equation modeling (SEM) or dynamic causal modeling (DCM) as well as Granger causality (which is sometimes considered as a functional connectivity approach).
- What is real-time fMRI?
Real-time fMRI refers to an approach within which the fMRI data are analyzed online as they are collected. This is made possible through the use of faster imaging sequences and appropriate computer analysis algorithms. The data analyzed in this fashion can be fed back to the participant who can learn to use this information for regulating the activity of some brain regions. Such biofeedback approach can be applied to patient populations, e.g., chronic pain patients who may learn to decrease the subjective feeling of pain by decreasing the BOLD signal within the targeted brain regions (deCharms et al., 2005).
Reporting an fMRI study
- Are there any guidelines on how to report procedures and results from an fMRI study?
Yes, there are quite a few guidelines on how to report an fMRI study. They are summarized as Standards here. However, although these standards are quite useful in providing guidance on how to report scanning and analysis parameters as well as the identified activations, it may sometimes still be quite difficult to be very straighforward and unambiguous in reporting (e.g., see potential problems with Reporting activations in Tables).
Combining fMRI with other methods
- What methods can fMRI be combined with?
fMRI can currently be combined with several other methods. Generally, it is very customary to record behavioral measures (participants’ responses) simultaneously with fMRI data. In contrast, recording physiological measures is performed in a somewhat lesser degree, although some of them are rather easy to measure, e.g., heart rate (actually, this is often done, but only for monitoring the participant during scanning), respiration or skin conductance. As for other methods which have been successively combined with fMRI, most prominent examples include eye tracking and EEG. Recently, same attempts have been done with TMS.
Although combining fMRI with other methods is very appealing and can be done quite successfully, this is still not a trivial task. Devices for recording the complementary data set (e.g., electrodes, amplifiers and cables for EEG) can only be used if they are made from a selected set of fMRI-compatible materials (because putting metal devices in the scanner can obviously be very dangerous). Even then, considerable artefacts can be expected in both recorded data sets, although not necessarily in the same degree. For example, when combining fMRI and EEG, data quality of fMRI is usually not very compromised, while the recorded EEG signal is dominated by the artefacts induced by the scanning environment.
Aguirre GK, Zarahn E, D'Esposito M (1998) The variability of human, BOLD hemodynamic responses. Neuroimage 8:360-369.
Bandettini, PA (2006) Functional MRI. Methods in Mind (Eds. C Senior, T Russell, M Gazzaniga). MIT Press.
deCharms RC, Maeda F, Glover GH, Ludlow D, Pauly JM, Soneji D, Gabrieli JD, Mackey SC (2005) Control over brain activation and pain learned by using real-time functional MRI. Proc Natl Acad Sci U S A 102:18626-18631.
Hasson U, Nir Y, Levy I, Fuhrmann G, Malach R (2004) Intersubject synchronization of cortical activity during natural vision. Science 303:1634-1640.
Horwitz B, Friston KJ, Taylor JG (2000) Neural modeling and functional brain imaging: an overview. Neural Netw 13:829-846.
Logothetis NK, Pauls J, Augath M, Trinath T, Oeltermann A (2001) Neurophysiological investigation of the basis of the fMRI signal. Nature 412:150-157.
Logothetis NK, Wandell BA (2004) Interpreting the BOLD signal. Annu Rev Physiol 66:735-769.
Menon RS (2001) Imaging function in the working brain with fMRI. Curr Opin Neurobiol 11:630-636.
Ogawa S, Lee TM, Kay AR, Tank DW (1990) Brain Magnetic-Resonance-Imaging with Contrast Dependent on Blood Oxygenation. Proceedings of the National Academy of Sciences of the United States of America 87:9868-9872.
Savoy, R.L., Bandettini, P.A., O’Craven, K.M., Kwong, K.K., Davis, T.L., Baker, J.R., Weisskoff, R.M., Rosen, B.R. (1995). Pushing the temporal resolution of fMRI: Studies of very brief visual stimuli, onset variability and asynchrony, and stimulus-correlated changes in noise. Proceedings of the Society of Magnetic Resonance Abstracts, 1, 450.
Visscher KM, Miezin FM, Kelly JE, Buckner RL, Donaldson DI, McAvoy MP, Bhalodia VM, Petersen SE (2003) Mixed blocked/event-related designs separate transient and sustained activity in fMRI. Neuroimage 19:1694-1708.