Pictures of Blood

A functional MRI scanner does not detect neural activity. It detects a difference in the magnetic properties of oxygenated and deoxygenated haemoglobin.

When a region of the brain becomes more active, it uses more oxygen, and the vasculature overcompensates by delivering more oxygenated blood than is consumed, so the local ratio shifts. That shift is the BOLD signal, and it is what the pictures are made of.

Three limitations follow from this and they are not technicalities.

It is slow. The blood flow response peaks around four to six seconds after the neural activity that prompted it, while neurons operate in milliseconds. The temporal resolution is thousands of times coarser than the process being studied.

It is coarse in space. A voxel is typically a couple of millimetres on a side, containing on the order of hundreds of thousands of neurons doing many different things. A region "lighting up" means a small average change across that population.

And it is indirect. The relationship between neural activity and blood flow is reliable enough to be useful and is a physiological chain with its own dynamics, which can be altered by caffeine, medication, age and disease.

None of this makes fMRI a bad tool; it has transformed what can be studied in a living human brain. It does mean the pictures are not photographs of thinking, and the gap between what they measure and what they are described as showing is where most of the trouble lives.

The commonest error has a name, and once you know it you see it everywhere.

Reverse inference is concluding that a mental process occurred because a region associated with it was active. The insula is active during disgust. It is also active during pain, empathy, uncertainty, craving, interoception and being rather warm. Observing insula activity therefore tells you very little about whether a subject felt disgust.

The inference would be valid if regions were selective, if this area did this one thing. Overwhelmingly they are not. Most regions participate in many processes, so activity is weak evidence about which process is occurring, and the strength of the evidence depends on a selectivity that is rarely quantified and usually low.

A large proportion of popular neuroscience is exactly this error, performed confidently.

The statistical problem is separate and was made famous by a fish.

A brain scan produces tens of thousands of voxels, each tested separately for a difference between conditions. At a five per cent threshold, thousands of them will show a difference by chance alone. Corrections for multiple comparisons exist and are essential, and in the mid-2000s a substantial fraction of published studies did not apply them adequately.

In 2009 Craig Bennett's team put a dead Atlantic salmon in a scanner, showed it photographs of humans in social situations, asked it to determine what emotion each person was experiencing, and analysed the data without correction. They found a significant cluster of activity in the salmon's brain cavity.

The poster won an Ig Nobel and did more to normalise proper correction than any number of methodological papers, because it was impossible to argue with and impossible to forget.

There is also a finding about the audience that everybody working with these images should know.

Studies have found that adding a brain image, or merely including neuroscientific language, makes a weak explanation more convincing to non-experts. The effect has been debated and refined since the first reports, and it is real. Something about a picture of a brain reads as physical evidence in a way that a bar chart of the same data does not.

What is actually being shown is a statistical map, thresholded at some level, overlaid on a structural template, frequently an average brain rather than any participant's, with the coloured regions indicating where a contrast exceeded a criterion. The colours are not activity levels; they are test statistics.

The complementary technology is worth knowing because its trade-offs are precisely inverted. EEG records voltage fluctuations at the scalp produced by synchronised activity across large populations of neurons. Its temporal resolution is milliseconds (it sees the process as it happens) and its spatial resolution is poor, because the skull smears the signal and many different arrangements of sources could produce the same pattern at the surface.

So fMRI tells you roughly where with poor timing, and EEG tells you exactly when with poor location. Neither tells you what, and combining them does not conjure that either.

Which gives a reasonable rule for reading any headline about a brain region. The underlying result is usually a small average difference in a signal from blood flow, in a region that does many things, across a modest number of participants. That may be a perfectly good finding. The claim about localisation and specificity is nearly always added afterwards, by somebody who did not run the study.