1.
Department of Physiology, Development and
Neuroscience,
2.
Department of
Psychiatry and Behavioral Sciences,
3.
Department of Computer and Information Sciences,
4.
Centre for Cognitive Science,
2.1 General Taxonomy of Experimental Design
2.2 Block and Event-related Designs
3.
PIPELINE OF fMRI DATA ANALYSIS
4.1 Experiment Stimulus Delivery
4.5 Databases and Meta Analysis
Neuroimaging
in vivo is becoming popular from the last
two decades. The primary quest of neuroimaging is to better-understanding the
functions of various brain areas pertaining to various cognitive processes of
interest. Though there are several neuroimaging techniques available currently,
the functional Magnetic Resonance Imaging (fMRI) is playing an important role
in the field of Imaging Neuroscience. In this paper an introduction to fMRI,
the issues related to experimental design and analysis will be presented. This
paper also discusses some of the neuroinformatics tools available for fMRI
research.
Key Words:
fMRI, Experimental design, Statistical Parametric Mapping
The functional Magnetic
Resonance Imaging (fMRI) is a powerful imaging tool that can be used to perform
brain activation studies non-invasively in vivo while subjects are engaged in
meaningful behavioural tasks. The resulting activation of the brain indirectly
depends on blood-oxygen-level-dependent (BOLD) signal (Ogawa et al., 1990).
Before the emergence of fMRI,
radioiostope based techniques, such as positron emission tomography (PET) which
measures regional cerebral blood flow (rCBF), were widely used for mapping the
brain function. However, these techniques are invasive and have a low spatial
and temporal resolution. Electroencephalography (EEG) which records the
electrical activity of nerve cells in the human brain by measuring electrical
potential on the scalp is also another widely used experimental technique in
neuroscience. Magnetoencephalography (MEG) is another non-invasive technique
which is becoming popular now-a-days. MEG measures the weak magnetic fields
generated above the scalp by current flow in the brain. This technique directly
measures the neuronal activity. EEG and MEG techniques though they probe brain
activity at very finer temporal resolutions, their spatial resolutions are very
poor.
The fMRI provides a
non-invasive method to access indirectly neuronal activity in the brain, by
measuring the haemodynamic metabolic signal i.e., blood-oxygen-level-dependant
(BOLD) signal. Effects of blood oxygen on the apparent transverse relaxation
time (T2*) were reported by Ogawa and colleagues (Ogawa et al., 1992).
Increased neuronal activity in a brain area leads to an increase in localized
cerebral blood flow, blood volume, and blood oxygenation. The BOLD fMRI
techniques are designed to measure primarily, changes in the inhomogeneity of
the magnetic field that result from changes in blood oxygenation.
Deoxyhaemoglobin is paramagnetic and introduces an inhomogeneity into the
nearby magnetic field, while oxyhaemoglobin is weakly diamagnetic and has
little effect. Hence, a decrease in deoxyhaemoglobin would cause an increase in
image intensity [refer to Hornak (2002) for more details of fMRI Physics]. fMRI
is well suited to measuring the dynamic changes in brain activity induced by
tasks that involve learning as it provides a reasonable spatial and temporal
resolution compared to other neuroimaging techniques such as positron emission
tomography [refer to Cohen and Bookheimer (1994); Volkow et al. (1997)].
After more than a decade of
fMRI research, there is still much to learn about how neuronal activity,
haemodynamics and fMRI signals are interrelated (Heeger and Ress, 2002). A
recent review of Ugurbil et al. (2003) suggested the possibility of obtaining
spatially accurate and quantitative data on brain function from magnetic
resonance technologies. Recently Chein and Schneider (2003); Culham (2005)
pointed out that there is a growing scientific and clinical community using
fMRI and the neuroimaging publications continue to increase exponentially.
Developing successful fMRI
experiments requires careful attention to experimental design, data acquisition
techniques, and data analysis. The experimental design is at the heart of any
cognitive neuroscience investigation. In this paper we present a brief review
of various issues related to the experimental design.
An fMRI
experiment to test a given psychological hypothesis must be designed within the
constraints of the temporal characteristics of the fMRI BOLD signal and of the
various confounding effects to which fMRI signal is susceptible. The most
important consideration is the actual design of the experiment. In conducting a
hypothesis-based experiment, we wish to be able to attribute any observed
effects to experimentally manipulated conditions. This can be guaranteed only
if conditions are randomly allocated to a presentation order for each subject
in a sensible manner. Further, this randomization should be appropriately
balanced, both across and within subjects. With such random allocation of
conditions, any unexpected effects are randomly scattered between the
conditions, and therefore do not affect the designed effects.
2.1
General Taxonomy of Experimental Design
Overall, designs can be
classified into three types i.e., categorical, factorial or parametric
(Friston, 1997). The categorical designs assume that the cognitive processes
can be dissected into sub-cognitive processes. That is one can remove and add
different cognitive processes by the assumption of pure insertion. The
categorical designs are further divided into subtraction type or conjunction
type. Cognitive subtraction designs are used to test the hypothesis pertaining
to activation in one task as compared to that in another task considering the
fact that the neural structures supporting cognitive and behavioural processes
combine in a simple additive manner. Whereas in the cognitive conjunctions type
designs, several hypotheses are tested, asking whether all the activations in a
series of task pairs, are jointly significant. Factorial designs involve
combining two or more factors within a task and looking at the effect of one
factor on the response to other factor. In parametric designs, rather than
assuming that the cognitive processes are composed of different cognitive
components, they are considered as belonging to different psychological
dimensions. The systematic changes in the brain responses according to some
performance attributes of task can be investigated in parametric designs. In
parametric designs one can also look at the linear and non-linear types of
relations to be determined emperically.
2.2
Block and Event-related Designs
The experimental designs are
broadly classified into two classes i.e., blocked (epoch) designs and
event-related designs. The properties of haemodynamic response function (i.e.,
the transfer function mapping neuronal activity onto BOLD signal) play an
important role in the design of experiments (Friston, 1998). Typically the task
condition in the blocked designs is performed for an extended period that is
more than the haemodynamic response (HR) time. Some of the advantages of
blocked designs are that they normally give reasonably significant areas of
activation and they are comparatively easy to analyze. On the other hand, the
event-related designs (sometimes called trial-based or single-trial designs)
aim to characterize transient changes in fMRI signal that emanate as the
consequence of individual trials either separated in time or spaced closely
together in time (Culham, 2005). Some of the advantages of using event-related
designs are: (1) allow random intermixing of the trials, (2) they are useful in
characterizing the temporal dynamics of brain activation, (3) allow separation
of sub-processes within multi-componential trials, (4) may facilitate
separation of HR signals from artifactual events etc. Thus the blocked designs
are special case of event-related designs and each has its own advantages and
disadvantages. In a recent investigation by Friston's group (Mechelli et al.,
2003), comparison between these two design
methodologies was presented taking a case study. Selection of experimental
design should be based on the particular research question of interest.
Prior to statistical analysis,
functional MRI data needs to be preprocessed so that images are registered to
each other (within and between subjects). These steps typically involve correction
for head movements, and normalizing to a stereotaxic space using a standard
brain template. In functional imaging, the signal changes corresponding to any
haemodynamic response can be small compared to the signal changes that can
result from subject motion. So, prior to performing the statistical tests, it
is important that the images are as closely aligned as possible. This step is
referred to as realignment or motion correction process. In addition to image
registration within subjects during the realignment process, images also need
to be registerred across subjects. This is done in order to account for
variation in brain sizes of different subjects and is achieved by normalization
or warping into a standard template space. The Talairach and Tournoux (1998)
template or a more representative template of the population provided by
Montreal Neurological Institute is often used for this purpose. Normalization
would also help standardization of reporting the co-ordinates of the brain
space across different studies. The fMRI images collected during the course of
an experiment are of low resolution compared to a structural image that clearly
identifies the structural properties of the brain. Prior to normalization, the
structural image and functional images are coregistered with each other. This
step finds the transformation that maps the structural image into the space of
the functional images. Now, the structural image can be used to find the
normalization parameters to achieve a precise spatial normalization. The
normalized functional images can then be spatially smoothed. The purpose of
spatial smoothing is to cope with functional anatomical variability that is not
compensated by spatial normalization and to improve the signal to noise ratio.
A Gaussian kernel (typically with an isotropic 6 to 8 mm full width at half
maximum) is used during the process of smoothing for validating any Gaussianity
assumptions in the statistical analysis fMRI data.
There are broadly two ways of analyzing
fMRI data: 1) model driven and 2) data driven. Model driven methods depend on
some hypothesis of the data that are usually tested within the framework of a
general linear model (GLM). To measure the magnitude of the BOLD signal that is
task-specific, neuroimaging data at each voxel are modeled as a linear
combination of explanatory variables plus a residual error term. The
explanatory variables model the design of the experiment, and it is possible to
incorporate noise due to measurement such as head movements in the form of a
design matrix. Other noise components include physiological factors such as
respiration and heart beating etc. As the BOLD signal does not measure the
neuronal activity, the experimental effects of interest are convolved with a
canonical haemodynamic response function. In order to account for variations in
the HRF, its temporal and dispersion derivatives could
be included in the design matrix. Thus each voxel is analyzed separately to
generate a test statistic (parameterized value) in a massively univariate
fashion. The end result is a parametric map that condenses information from a
number of individual scans into a single image volume that can be more easily
viewed and interpreted.
An alternative approach is to
use model-free methods such as the Independent Component Analysis (ICA).
cardiac and
respiratory effects, subject movements, and noise. Model-free approaches such
as
4.1
Experiment Stimulus Delivery
Stimulus delivery is an
integral part of conducting an fMRI experiment. While subjects lie supine in an
fMRI scanner, specific stimuli can be delivered through a
computer controlled software. Presentation (http://www.neurobs.com/presentation)
is a high-precision commercial program for stimulus delivery and experimental
control for behavioral and physiological experiments. E-Prime (http://www.pstnet.com/products/e-prime/)
is another commercial product for experiment generation and millisecond
precision data collection. Matlab (http://www.mathworks.com/)
is a high-level interpreted language available commercially with extensive
support for numerical calculations. The Psychophysics Toolbox (http://psychtoolbox.org/wikka.php?wakka=HomePage)
provides a Matlab interface to the computer’s hardware. Very useful for vision research, the
PsychoPhysics toolbox, also provides synchronization with vertical retrace of
the display, support millisecond timing, sound, keyboard, and the serial
port. Cogent 2000 (http://www.vislab.ucl.ac.uk/Cogent/)
is a Matlab toolbox available for researchers as a freeware. Cogent 2000 is
useful for presenting stimuli and recording responses with precise timing. It
incorporates Synchronization with fMRI scanner and Cogent Graphics provides
additional utilities for the manipulation of sound, keyboard, mouse, joystick,
serial port, parallel port, subject responses and physiological monitoring
hardware.
DICOM (Digital Imaging and
Communications in Medicine) is the most common standard for medical images (http://dicom.nema.org), such as Computed
Tomography (CT), MRI, and Ultrasound. A single DICOM file contains both a
header (which stores information about the patient's name, the type of scan,
image dimensions, etc), as well as all of the image data (which can contain
information in three dimensions). This is different from the popular Analyze
format (http://www.mayo.edu/bir/PDF/ANALYZE75.pdf),
which stores the image data in one file and the header data in another file.
Recently, the Neuroimaging Informatics Technology Initiative (NIfTI) has become
handy for the purpose of having a coordinated image format that would be
compatible with a number of Neuroinformatics tools (http://nifti.nimh.nih.gov). A number of
software tools allow conversion from one data format to another. For a detailed
description, refer to MRIcro software (for medical image visualization) guide
by Chris Rorden (http://www.sph.sc.edu/comd/rorden/mricro.html).
A number of popular software
tools such as AFNI (Cox, 1996), FSL (FMRIB Software Library – Smith et al.,
2004) and SPM (Statistical Parametric Mapping – refer to edited book by Friston
et al., 2006) are available freely for the neuroimaging community. Dedicated
softwares for preprocessing such as AIR (Automated Image Registration http://www.loni.ucla.edu/Software/Software_Detail.jsp?software_id=8)
are available, although many software tools for fMRI data analysis include
preprocessing routines. AFNI (http://afni.nimh.nih.gov/afni)
is a set of C programs for processing, analyzing, and displaying fMRI data and
runs under Unix, Linux and MacOS environments. The SPM
software is a suite of MATLAB functions and subroutines. The current release of
SPM (http://www.fil.ion.ucl.ac.uk/spm)
is designed for the analysis of fMRI, PET, SPECT, EEG and MEG data. FSL (http://www.fmrib.ox.ac.uk/fsl)
contains statistical tools for FMRI, MRI and DTI (Diffusion Tensor Imaging)
data. The set of tools available in FSL include both model-based (FEAT - FMRI
Expert Analysis Tool) and model-free (MELODIC - Multivariate Exploratory Linear
Optimized Decomposition into Independent Components) analysis methods. Group
Interpretation of the outcome
of an fMRI experiment can be done by projecting the activation maps on to the
structure of the brain either as an array of 2D slices or onto the surface of
the brain. Softwares like MRIcro (http://www.sph.sc.edu/comd/rorden/mricro.html)
and mri3dX (http://www.aston.ac.uk/lhs/research/groups/nrg/mri3dx/index.jsp)
allow slice based as well as 3D viewing. The location of activation can also be reported on
the basis of its coordinates in the stereotaxic space such as the Talairach and
Tournoux. The Talairach daemon (http://ric.uthscsa.edu/projects/talairachdaemon.html)
has a digitized database of the atlas by Talairach and Tournoux (1998) and a
Java based interface that can read text files with x, y, z coordinates and
return the location of activation in terms of 5 level heirarchy of label names
including the hemisphere, lobe, gyrus, gray/white matter and the brodmann area.
The Automated Anatomic Labelling (AAL - http://www.cyceron.fr/freeware/) is a extension for SPM and consists of the parcellation done on
a high Resolution MNI single subject brain. An AAL template is also available
with MRIcro. SUMA (http://afni.nimh.nih.gov/afni/suma/)
is a program that adds cortical surface based functional imaging analysis to
the AFNI suite of programs. It allows viewing 3D cortical surface models, and
mapping volumetric data onto them.
4.5
Databases and
The fMRI data centre (http://www.fmridc.org/f/fmridc)
provides complete data sets (raw, functional and structural MRI) from many peer
reviewed fMRI studies. The journal of cognitive neuroscience requires all
papers accepted for publication to deposit the data to the fMRIDC. To encourage
scientific exchange, a new perspectives in fMRI
Research award has been initiated for novel utilization of any of the datasets
in the fMRIDC. BrainMap (http://www.brainmap.org/)
is an online database of published functional neuroimaging experiments with
coordinate-based (Talairach) activation locations. It is a tool to rapidly
retrieve and understand studies in specific research domains, such as language,
memory, attention, reasoning, emotion, and perception, and to perform
meta-analyses of like studies. The number of fMRI papers grows far too fast for
anyone to read them all. AMAT (A Meta Analysis Toolbox – http://www.dartmouth.edu/~antonia/amat.html)
is a matlab program which allows searching through the coordinates reported in
lots of fMRI papers.. The Brede database (http://hendrix.imm.dtu.dk/services/jerne/brede/)
also allows coordinate based searching.
Neuroimaging with functional
MRI has attracted researchers to assess the function of human brain in vivo. This article has presented the
main issues in experimental design and generic approaches for data analysis. We
have given pointers to some of the useful neuroinformatics tools that have
gained popularity. Newer methods of data analysis and interpretations have been
developed more recently. For example, multivariate techniques, bayesian analysis (Friston et
al., 2002) for statistical analysis of fmri data and connectivity among various
brain regions (Friston et al., 2003). There has been a growing interest
among usefulness of functional MRI to be able to predict the cognitive state a
person is at a given instantaneous time point (Mitchell et al., 2004). This
would enable applications like “brain reading” (Haynes and Rees, 2006) and “lie
detection” (Davatzikos et al., 2005). The functional MRI has now gained an
increased interest for use in clinical research (Partain, 2006). It appears
that fMRI as a tool has many more exciting applications impending in near
future.
Biswal,
B.B., Ulmer, J.L., 1999. Blind source separation of multiple signal sources of
fMRI data sets using independent component analysis. J. Comput. Assist. Tomogr. 23, 265– 271.
Chein, J. M. and Schneider, W. (2003). Designing
effective fMRI experiments. In Grafman,
J. and Robertson,
Cohen,
M. S. and Bookheimer, S. Y. (1994). Functional
magnetic resonance imaging. Trends in Neurosciences, 17:268–277.
Cox, R.
W. (1996) AFNI – Software for Analysis
and Visualization of Functional Magnetic Resonance Neuroimages.
Computers and Biomedical Research 29:162-173.
Culham,
J. C. (2005). Functional neuroimaging: Experimental design and analysis. In Cabeza, R. and Kingstone, A., editors, Handbook of Functional
Neuroimaging of Cognition, 2nd edition. MIT Press,
Davatzikos
C., Ruparel K., Fan Y., Shen D. G., Acharyya M., Loughead J. W., Gur R. C., and
Langleben D. D. (2005). Classifying spatial patterns of brain activity with
machine learning methods: application to lie
detection. NeuroImage, 28:663–668.
Friston,
K. J. (1997). Imaging cognitive anatomy. Trends in
Cognitive Sciences, 1:21–27.
Friston,
K. (1998). Imaging neuroscience: principles or maps? Proc. Natl. Acad. Sci.
USA, 95:796–802.
Friston
K.J., Glaser D.E., Henson R.N.A., Kiebel S., Phillips C., and Ashburner J.
(2002). Classical and Bayesian Inference in Neuroimaging:
Applications. NeuroImage, 16:484-512.
Friston,
K. J., Harrison, L., and Penny, W. (2003). Dynamic
causal modelling. NeuroImage, 19:1273–1302.
Friston,
K. , Ashburner, J., Kiebel, S., Nichols, T., and
Penny, W., Eds, (2006) Statistical Parametric Mapping: The analysis of
functional brain images. Elsevier,
Haynes,
J-D., and Rees G. (2006). Decoding
mental states from brain activity in humans. Nature Reviews Neuroscience
7: 523–534.
Heeger,
D. J. and Ress, D. (2002). What does fMRI tell us about
neural activity? Nature Reviews, 3:142–150.
Hornak,
J. P. (2002). The basics of MRI. Online book at: http://www.cis.rit.edu/htbooks/mri/,
Makeig,
S., Jung, T.P.,
McKeown,
M.J., Jung, T.P., Makeig, S., Brown, G., Kindermann, S.S., Lee, T.W.,
Sejnowski, T.J., 1998. Spatially independent activity patterns in functional
MRI data during the stroop color-naming task. Proc. Natl. Acad. Sci. U. S. A.
95, 803–810.
Mechelli,
A., Henson, R. N. A., Price, C. J., and Friston, K. J. (2003).
Comparing event-related and epoch analysis in blocked design fMRI. NeuroImage,
18:806–810.
Mitchell,
T., Hutchinson, R., Niculescu, R., Pereira, F., Wang, X., Just M., and Newman,
S. (2004). Learning to decode cognitive states from brain
images. Machine Learning, 57:145–175.
Ogawa,
S., Lee, T. M., Kay, A. R., and Tank, D. W. (1990). Brain magnetic resonance imaging with contrast dependent on blood
oxygenation. Proceedings of the National
Ogawa,
S., Tank, D. W., Menon, R., Ellermann, J. M., Kim, S., Merkle, H., and Ugurbil,
K. (1992). Intrinsic signal changes accompanying sensory
stimulation: Functional brain mapping with magnetic resonance imaging.
Proceedings of the National
Partain,
C. L. (2006). JMRI special issue: Clinical potential of brain mapping using
MRI, Journal of Magnetic Resonance Imaging, 23:785–786
Smith,
S.M., Jenkinson, M., Woolrich, M.W., Beckmann, C.F., Behrens, T.E.J.,
Johansen-Berg, H., Bannister, P.R., De Luca, M., Drobnjak, I., Flitney, D.E.,
Niazy, R., Saunders, J., Vickers, J., Zhang, Y., De Stefano, N., Brady, J.M.,
and Matthews, P.M. (2004). Advances in functional and structural MR image analysis
and implementation as FSL. NeuroImage, 23(S1):208-219
Talairach
P, Tournoux J. (1988). A Stereotactic Coplanar Atlas of the
Human Brain.
Ugurbil,
K., Toth, L., and Kim, D. S. (2003). How accurate is magnetic
resonance imaging of brain function? Trends in Neuroscience, 26:108–114.
Volkow,
N. D., Rosen, B., and Farde, L. (1997). Imaging
the living human brain: Magnetic resonance imaging and positron emission
tomography. Proceedings of National