Magnetic resonance imaging (MRI) has played a
major role in the investigation of brain structure,
function development, and pathologies. MRI experiments
on the human brain can be divided into two categories:
structural experiments, which rely on the biophysical
properties of brain tissue, and functional experiments,
which are sensitive to temporarily changing neuron
activity. The interrogation and analysis of MRI data is
exceptionally challenging in terms of the software
required to produce three-dimensional images from the
An analysis of the field of Neuroscience and Behavior in
Essential Science IndicatorsSMfrom
the past decade has highlighted the publication record of Dr. Tim Behrens.
His record in the database includes 1,500 total citations for 35 papers
published in this field alone from January 1, 1999 to June 30,
Dr. Behrens is a
Research Lecturer at the University of Oxford, working for
the Centre for Functional MRI of the Brain at the John
Radcliffe Hospital. In this interview,
European correspondent Dr. Simon Mitton examines the key
contributions to computational neuroscience made by Dr.
Behrens and his colleagues.
Your Masters degree is in engineering, so
how did you come to be working in neuroscience?
At Oxford I studied information engineering and machine learning for my
MEng degree. I was interested in doctoral research, but could not settle on
an area. Then I became attracted to brain imaging by an ebullient person,
Professor Sir Michael Brady, who is a leading figure in the disciplines of
artificial intelligence, medical image analysis, and robotics. After I had
made a literature review it struck me that diffusion imaging was an
exciting potential tool for mapping the connections in the brain.
I adopted an information engineering approach to see if we could make it
work. The question I posed was this: can diffusion MRI measurements be used
to infer something about the connections between different regions of the
brain? A number of groups were working on it at the time, so it was an
How do you like to describe your present research
Using probabilistic diffusion tractography (Behrens
et al. MRM 2003) we were able to segment the human
thalamus on the basis of its connectivity to the
cortex (Behrens et al. Nat Neurosci 2003).
View other probabilistic
diffusion tractography figures with
descriptions at the Web site of the University
of Oxford's FMRIB Centre, Department of
I have two main research interests. The first is in understanding
anatomical brain connectivity, which you can picture as the paths of the
wires that connect different regions of the brain. My second interest is to
understand learning and decision-making in the human and macaque brains
from a computational perspective.
Your most-cited papers are on the application of
diffusion imaging for structural and functional mapping. What are the
principles underlying this technique?
The human brain is made up of neurons, and they send signals to other
neurons via their axons. Axons act like the wiring in a phone network.
Obviously it is of great importance to understand how different parts of
the brain are interconnected because you cannot understand what a region of
the brain does until you know what information it has access to, and what
influence it might have on other brain regions. The challenge, of course,
has always been to develop imaging techniques that can safely be used on a
About 20 years ago, researchers at the National Institutes of Health made a
breakthrough with developing diffusion MRI to image brain structure in
stroke patients. Diffusion MRI exploits the fact that the neural axons are
insulated with an outer sheaf of myelin, which is fat. Water finds it
difficult to penetrate myelin, so if you can imagine trying to measure
diffusion in an axon, you would find that water molecules diffuse
preferentially along an axon rather than across it. So you can tell the
direction of an axon by looking at the direction of local diffusion.
How have you contributed to the development of diffusion
I have designed techniques that will trace diffusion paths, in vivo, to let
us measure the connections between brain regions. I joined the field
through the technical aspects of making diffusion MRI work. My top two
papers (Behrens TEJ, et al., "Non-invasive mapping of connections
between human thalamus and cortex using diffusion imaging," Nat.
Neurosci. 6: 750-7, 2003; Smith SM et al., "Advances in
functional and structural MR image analysis and implementation as FSL,"
NeuroImage 23: S208-19, 2004) show some of the very first results
from this field, demonstrating that you can map out the connections between
different brain regions.
As an engineer trying to improve the technology of brain imaging, I have
needed to work alongside enthusiastic neuroscientists who can tell me
whether my measurements and the pathways they represented look reasonable.
That’s how I came to collaborate with Heidi Johansen-Berg, who became
interested in the technique I was developing. We did a lot of work on
mapping the anatomy of the human brain by comparison to what is known in
monkeys, with me providing the computational expertise and Heidi the
anatomical expertise ("Changes in connectivity profiles define functionally
distinct regions in human medial frontal cortex," Proc. Natl. Acad.
Sci. USA 101: 13335-40, 2004; "Functional-anatomical validation and
individual variation of diffusion tractography-based segmentation of the
human thalamus," Cereb. Cortex 16: 1418-30, 2005).
Ever since then I have been working on the boundary of methods development
and neuroscience. Almost everything I do involves methods that I have had
to develop in some way. But over the last 10 years I have become a
neuroscientist as well as an engineer.
When I began, the field needed moving out of the realm of simply looking at
big white matter bundles, which are easy to measure, into the more subtle
question of asking which grey-matter regions are connected to each other.
This is important, because it is the grey-matter regions that contain the
cell bodies where computations are performed. To move beyond that I
developed algorithms that made much better use of the raw data, so it
became possible to ask a whole host of new questions.
Can you give me an example?
"The challenge, of course, has
always been to develop imaging techniques
that can safely be used on a live
My quantitative work has included, for example, calculating the probability
that two regions of the brain are connected. In this way it became possible
to trace the pathway from grey matter to grey matter. That’s all
covered in the 2003 Nature Neuroscience paper on the non-invasive
mapping of thalamic connections. That’s the first paper that showed
how to trace from grey matter to grey matter, with quantitative mapping.
The paper covers the techniques we developed, but it also demonstrates that
the architecture of the thalamus in the human brain is similar to that of
the macaque monkey.
The citations to your papers indicate worldwide interest
in your techniques. How did that arise?
The key technical papers are the three published in the journal
NeuroImage: the 2004 Smith SM, et al. paper, as well as
Behrens TEJ, et al., "Probabilistic diffusion tractography with
multiple fibre orientations: What can we gain?" (34: 144-55, 2007); and
Smith SM, et al., "Tract-based spatial statistics: Voxelwise
analysis of multi-subject diffusion data," (31: 1487-1505, 2006). These are
technical accounts of new techniques for imaging different regions of the
The 2004 paper with Steve Smith describes a whole set of tools that are
used worldwide. I should emphasize that these papers are from the large
functional imaging group in Oxford headed by Steve Smith, which consists of
nearly 20 scientists, most of whom work on tools, only one of which is the
method I have described. We have produced a software library that is freely
used by 5,000 labs worldwide to analyze brain data. One of the tools is the
diffusion imaging I have been describing. Other tools look at functional
imaging, structural imaging; they basically let you interrogate and
understand your brain imaging data.
Can you give me examples of the clinical
The software from our imaging group is used to map out connections in
different parts of the brain. The clinical applications include looking at
how connections might change in different diseases. A number of citations
are by clinical researchers who, having found activations in the thalamus,
then want to know to which cortical region it connects. We have gone on to
develop new techniques for understanding the influence that anatomical
connectivity patterns have on functional specialization in grey matter.
This is taken up in the 2004 PNAS paper, which presents the
information in atlas format. So if a patient has a lesion in the thalamus
it is possible to conclude which region of the cortex will be affected.
As another example, there’s a new type of treatment for motor
disorders that involves implanting a stimulator into particular
sub-cortical nuclei, but it is very hard to identify where the nuclei are
located in an individual brain. Our technique lets you take a big
sub-cortical structure and locate within it the target, for example, the
motor nucleus in the case of motor disorder. This may enable a neurosurgeon
to position the deep brain stimulator correctly.
Your second interest, learning and decision-making
research, has focused on understanding the computational algorithms
that are employed in human learning, and how they affect human
behavior. What aspects of learning interest you?
"If you need one, try this: "The
software from our imaging group is used to
map out connections in different parts of the
I work with Matthew Rushworth. We are trying to understand decision-making
from a more formal perspective than most researchers. I build models of how
people learn, and I use those models to predict what choices they will make
when presented with a decision, and also to predict activity in parts of
the brain. This is a new and exciting idea in neuroscience because we
combine computational neuroscience with recordings of brain activity.
I try to understand the algorithms the brain uses for learning. For
example, you can show that certain computational parameters are necessary
for particular patterns of behavior. You can then solve an inverse problem:
by witnessing a pattern of behavior you have evidence that the parameter is
coded somewhere in the brain. We look for brain activity that predicts or
correlates with how that hidden internal parameter in the computational
model is changing.
Together with Dr. Rushworth, I have concentrated on how the brain combines
recent information with our historical experiences. We show in our 2007
Nature Neuroscience paper, "'Learning the value of information in
an uncertain world" (Behrens TEJ, et al., 10:1214-21), that a key
computational parameter for performing this task optimally is coded in the
Anterior Cingulate Cortex (ACC) in the course of learning. Our 2006
Nature Neuroscience paper, "Optimal decision making and the
anterior cingulate cortex" (Kennerley SW, et al., 9:940-7) shows
that removal of this ACC region causes a specific deficit in this aspect of
In conclusion Dr. Behrens, could you tell me what you
are working on right now?
Recently we have been toying with the idea that we might be able to use
formal computational models of behavior to model social interactions
between individuals, and so we can find parallels between parameters that
are rather similar in two different domains: one where you are learning
about your own behavior, and the other where you are learning about someone
In the ACC, one location codes for learning about rewards, and an adjacent
location codes for the same computational parameter when learning about
someone else’s behavior. Based on which area has more activity we can
predict whether somebody is going to be more influenced by their own
interactions with the environment or more influenced by advice from other
people. We are using mathematical formalism to model social behavior in
terms of brain mechanisms.
Tim Behrens, MEng DPhil
Department of Experimental Psychology
University of Oxford
Oxford, United Kingdom