Brain Imaging Methods

The problem every method is trying to solve

Almost every other organ can be studied by opening it up, cutting a piece out, or watching it move. The brain resists all three. It is enclosed in the skull, which is opaque to visible light and hostile to most forms of access. Its functional units, the neurons and the synapses between them, are measured in micrometres. And the events that matter, the action potentials and synaptic potentials that constitute its work, are over in a few thousandths of a second.

So there is no method that simply "shows the brain thinking". Every technique makes a bargain, and the bargain is always struck along three axes at once.

Understanding a method therefore means understanding its bargain. It is not enough to know that fMRI "shows brain activity"; you have to know that it does so by watching blood, several seconds late, and that this single fact determines almost every question it can and cannot answer. What follows walks through the major methods with that discipline: what it measures, what it is genuinely good for, and where its hard limit lies.

Structural MRI: exquisite anatomy, silent about function

Magnetic resonance imaging exploits a property of hydrogen nuclei, which are abundant in the body because the body is largely water. Placed in a very strong magnetic field, these nuclei align with it. A radiofrequency pulse knocks them out of alignment, and as they settle back they emit a faint signal whose timing depends on the local chemical environment. Fat, water, grey matter and white matter relax at measurably different rates, so the machine can distinguish them. Magnetic gradients encode position, allowing the signal to be reconstructed into a three-dimensional image.

The result is the most detailed non-invasive picture of the living human brain that exists, resolving structures below a millimetre, with no ionising radiation. It is why a person with a suspected tumour, a suspected multiple sclerosis plaque, or a suspected stroke goes into an MRI scanner. It is also the basis of morphometric research, which measures cortical thickness, the volume of the hippocampus, or the integrity of white matter tracts.

Its limit is absolute and easily forgotten. Structural MRI is a photograph. It tells you the shape of the tissue and nothing whatsoever about what that tissue is doing. A brain that is thinking hard and a brain that is anaesthetised produce the same structural scan. Every claim about function that appears to come from an MRI machine actually comes from a different sequence run on the same hardware, which is the subject of the next section.

fMRI: why the most famous method measures blood, not thought

Functional MRI is the technique behind almost every brain image the public ever sees, and it is the most systematically misunderstood method in neuroscience. The misunderstanding has a single root: people assume it records neural activity. It does not. It records blood.

Here is the actual causal chain, which is worth following link by link.

1. Neurons work, and working costs energy

   A population of neurons becomes more active. Sustaining and restoring their electrical gradients consumes adenosine triphosphate, and regenerating that ATP consumes oxygen and glucose. This is covered in detail under brain energy and metabolism.

2. The local blood supply responds, and overshoots

   Through a process called neurovascular coupling, signals released by neurons and by astrocytes cause nearby arterioles to dilate, and local blood flow rises. Crucially, the flow increase exceeds what the tissue actually extracts. More oxygenated blood arrives than is consumed.

3. The magnetic properties of the blood change

   Haemoglobin carrying oxygen is diamagnetic; haemoglobin that has given up its oxygen is paramagnetic, meaning it distorts the surrounding magnetic field. Because the flow overshoot reduces the local proportion of deoxygenated haemoglobin, that field distortion decreases, and the MRI signal in that voxel rises slightly.

4. The scanner detects the change, several seconds late

   This is the blood-oxygen-level-dependent, or BOLD, response. It begins about a second or two after the neural event, peaks at roughly four to six seconds, and takes many seconds more to return to baseline. The neural event itself lasted milliseconds.

What follows from all of this is a precise account of the method's strengths and its ceiling. Spatially, fMRI is good: typical voxels are a few millimetres on a side, and it sees the whole brain including deep structures such as the basal ganglia and the thalamus that scalp electrical methods cannot reach. Temporally, it is poor and irreparably so: even a perfect scanner cannot beat the sluggishness of the blood response it depends on. And, most importantly of all, it is correlational. An fMRI experiment shows that a region's blood oxygenation changed while a person did a task. It cannot show that the region was doing the task, that it was necessary for the task, or that the task could not have been done without it. That question requires a different class of method entirely.

EEG and MEG: the only methods fast enough to watch the brain work

If fMRI is slow because it watches blood, the electromagnetic methods are fast because they watch the electricity itself.

The pairing is instructive. EEG and MEG measure the same underlying biophysical event and differ mainly in what the head does to the signal on its way out. That a change of physical medium, from electric to magnetic, buys you spatial precision at a cost of a shielded room and a cryogenic dewar is a fair illustration of how the trade-offs in this field actually work: they are not arbitrary, they follow from physics, and they cannot be talked away.

PET: the only method that can image a molecule

Positron emission tomography takes a wholly different approach. A molecule of biological interest is synthesised with one of its atoms replaced by a radioactive isotope that decays by emitting a positron. The tracer is injected into the bloodstream. Each emitted positron travels a short distance, meets an electron, and the pair annihilates, producing two gamma photons that fly apart in opposite directions. A ring of detectors around the head registers pairs of photons arriving simultaneously, and from many such coincidences the scanner reconstructs where in the brain the tracer accumulated.

What makes PET indispensable is not resolution. Its spatial resolution is worse than MRI and its temporal resolution is dismal, ranging from tens of seconds to minutes depending on the tracer. What makes it indispensable is chemical specificity. Choose your tracer and you choose what you image.

The costs are real and set the boundary on its use. PET involves ionising radiation, which restricts how often a person can be scanned and effectively excludes healthy children and repeated study in most healthy volunteers. Many tracers are made with short-lived isotopes and require an on-site cyclotron. And the temporal resolution is such that PET can tell you what the brain's chemistry looked like over a period, never what it did in a given second.

Causal methods: the difference between active and necessary

Everything discussed so far, with the exception of nothing at all, is correlational. It records what happens while something else happens. That is a genuine limitation, and it is the deepest one in the field, because the questions people actually care about are causal: is this region doing the job, or is it merely nearby while the job is done?

There are two ways to get a causal answer in a human being, and both work by intervening rather than observing.

The animal methods, and why they are the gold standard

Almost everything mechanistic that is known about how neurons compute was learned by methods that cannot be used on people, and it is worth being explicit about this, because it explains why so many confident popular claims about the human brain are extrapolations.

Single-unit recording means lowering a fine microelectrode into the brain until its tip is close enough to a single neuron to register that neuron's individual action potentials. The resolution is exquisite: you are watching one cell, spike by spike, in real time. This is how the visual cortex's orientation-selective cells were discovered, how place cells and grid cells were found in and around the hippocampus, and how much of what we know about how information is coded in firing rates was established. Multi-electrode arrays now record hundreds of cells at once. In humans it is possible only in rare clinical circumstances, chiefly in epilepsy patients who already have electrodes implanted for surgical planning.

Optogenetics goes one step further and is worth understanding properly, because it is the reason the last two decades of systems neuroscience look the way they do. Certain algae and bacteria make light-sensitive ion channels, proteins that open when struck by light of a particular colour. Using genetic techniques, these channels can be inserted into a chosen population of neurons in a living animal, and into that population only, targeted by cell type. An optical fibre then delivers light to the region. Because opening an ion channel changes the neuron's membrane potential, a flash of light can now switch that specific class of neurons on, or, with a different opsin, off, on a millisecond timescale, in a behaving animal.

The trade-off, stated plainly

The whole methodological landscape can be compressed into a single observation: nobody has a method that is simultaneously precise in space, precise in time, non-invasive and causal. Choose three at most, and usually two.

Read down that list and a strategy suggests itself, and it is the strategy good laboratories actually use: converge. Combine methods whose weaknesses do not overlap. Use fMRI to find out where, EEG or MEG to find out when, TMS or lesion evidence to find out whether it is necessary, and animal work to find out how the cells do it. A conclusion supported by one method is a hypothesis. A conclusion supported by four methods with independent failure modes is knowledge.

How brain imaging is misused

This is the section that matters most, and it is worth dwelling on. Neuroimaging is not a fringe field peddling nonsense; it is a serious and productive science. But its outputs are unusually easy to misread, partly because the images are so persuasive, and there are four specific failures that recur relentlessly, in the press and sometimes in the literature.

1. Nothing "lights up"

The phrase "the X centre lit up" is not merely loose language; it encodes a false picture of what happened. There is no light. There is no centre being switched on. The brain does not have quiet regions that activate when needed; the whole brain is metabolically busy at all times, which is exactly why it consumes about a fifth of the body's energy while you do nothing (see brain energy and metabolism). What an fMRI experiment produces is a difference: the signal in condition A minus the signal in condition B, tested voxel by voxel, with the voxels that cross a statistical threshold coloured in. The coloured blob is where the difference was reliable. The uncoloured brain is not switched off. It simply did not differ between the two conditions the experimenter happened to compare, and if the experimenter had compared different conditions, a different blob would have appeared.

2. Reverse inference

This is the central logical error of popular neuroscience, and it was named and dissected by Russell Poldrack in 2006. The reasoning runs: in this experiment, region R was active; in previous studies, region R has been associated with mental state M; therefore the participants were experiencing M.

Once you can see this error you will find it everywhere: the neuromarketing study that claims consumers "loved" a product because a reward-associated region was active; the courtroom claim that a defendant's scan reveals a lack of empathy; the headline claiming that a scan shows people are "addicted" to their phones because a dopamine-associated region responded. Each is a reverse inference, and each is a fallacy.

3. Multiple comparisons, and the dead salmon

A functional scan divides the brain into voxels, and a typical whole-brain analysis contains tens of thousands of them. The standard analysis tests each voxel separately for a difference between conditions. Now consider what a significance threshold of p < 0.05 means: it means that, when there is no real effect at all, a given test will nevertheless come out "significant" about five per cent of the time by chance. Run that test on forty thousand voxels with no real effect anywhere, and you should expect roughly two thousand voxels to pass. Some of them will be adjacent, and adjacent significant voxels form a cluster, and a cluster looks exactly like a finding.

This is not a hypothetical. Craig Bennett and colleagues made the point unforgettably by placing a dead Atlantic salmon in an fMRI scanner, showing it photographs of humans in social situations, and asking it to determine what emotion the people were experiencing. Analysing the data without correcting for multiple comparisons, they found a cluster of "significantly active" voxels in the dead fish's brain cavity. The fish was, they noted drily, not alive at the time of scanning. The work, presented in 2009 and published in 2010, was a deliberate demonstration, and its target was not fMRI but a sloppy practice within it: at the time, a substantial fraction of published papers were not correcting properly. The remedy is standard and well understood, involving corrections such as controlling the family-wise error rate or the false discovery rate, and it is now expected. But the salmon remains the clearest available illustration of why an uncorrected blob is worth nothing.

4. Small samples and low power

Neuroimaging is expensive. Scanner time costs hundreds of pounds an hour, which for many years pushed studies toward samples of fifteen or twenty participants. A small sample means low statistical power, and low power has two consequences that are widely misunderstood. The first is the obvious one: you will miss real effects. The second is worse and less intuitive: among the effects that do reach significance in an underpowered study, a higher proportion are false positives, and the true effects that survive are systematically overestimated in size, because only an unusually large sample estimate could have crossed the threshold. A literature built on small studies therefore accumulates findings that are too few, too large, and too often wrong, and it replicates poorly. The field has recognised this, and the response has been a move toward much larger samples, data sharing, and pre-registration of analysis plans, which is a genuine and creditable reform. But a great deal of the older literature, including much of what has filtered into popular knowledge, was produced under the old regime.

Myths about brain scans

How to read a neuroimaging claim

The practical value of everything above is that it gives you a short checklist. Applied to almost any brain-imaging story, it will tell you quickly whether there is anything there.

None of this is a reason to distrust neuroimaging. It is a reason to read it as the practitioners themselves do: as a set of powerful but sharply limited instruments, each of which answers one kind of question well and lies convincingly about the others.

Related topics