Functional MRI

Haemoglobin is diamagnetic when carrying oxygen and paramagnetic when not. That difference distorts the local magnetic field enough to change the MR signal, so a sequence sensitive to field inhomogeneity reports on blood oxygenation — the BOLD contrast. Ogawa’s observation was that this happens without any injected agent: the contrast is already in the blood.

Linus Pauling and Charles Coryell had reported the magnetic difference in 1936. Seiji Ogawa and colleagues at AT&T Bell Laboratories made it visible in 1990: in gradient-echo images of anaesthetised mice and rats at 7 and 8.4 T, blood vessels showed as dark lines when their deoxyhaemoglobin rose. That December they named the effect, showed it tracking blood oxygen, and offered it for PET-like measurements of brain activity.

From an injected agent to the blood’s own

The first functional MR maps of a human brain, published by John Belliveau and colleagues at Massachusetts General Hospital in November 1991, did use an injected agent. They tracked boluses of gadolinium through the brain with fast imaging, once at rest and once during visual stimulation, and found blood volume in primary visual cortex up by 32%. Each condition cost an injection.

By the time they appeared, the injection was unnecessary. On 9 May 1991, at the same hospital, Kenneth Kwong imaged visual cortex with a gradient-echo sequence while a volunteer’s goggles flickered on and off, and the cortex lit up on the first run. At the University of Minnesota, Ogawa and Kâmil Uğurbil had been attempting the same on a new 4 T scanner since early 1991 and had positive results by that August; the efforts were independent, and which came first has been disputed. Peter Bandettini and Eric Wong of the Medical College of Wisconsin saw Kwong’s images in a plenary lecture that August and had motor cortex activation of their own a month later.

All three groups published in June and July 1992. Kwong’s team, at 1.5 T, reported a 1.8% signal rise in the visual cortex of seven volunteers, with a time constant of about 4 s; Bandettini’s, 4.3% in motor cortex; Ogawa’s, at 4 T, 5–20%, shrinking as the echo time was cut from 40 ms to 8 ms — the signature of a change in $T_2^*$, the dephasing that deoxyhaemoglobin causes. Jens Frahm’s group in Göttingen followed later that year. Belliveau’s method lived on as a clinical measure of perfusion; for mapping function it was overtaken within months.

Echo-planar imaging, and the gradients behind it

The MR signal from a slice is the slice’s two-dimensional Fourier transform, sampled along a path set by the running integral of the gradient fields. Early line-scanning methods built an image a line at a time, and a 64 × 64 image took 10–20 minutes. Peter Mansfield’s echo-planar imaging, published in 1977, gathers the whole plane after one excitation, reversing one gradient back and forth while the other steps the path along, so that a single train of echoes makes an image in 20–50 ms. It has to finish before the signal decays, within a $T_2^*$ of a few tens of milliseconds, which is why the first fMRI images were 64 pixels square.

It took about a decade to become practical, because it asks a great deal of the gradient coil. Switched fast inside a superconducting magnet, the coil induces eddy currents in the cryostat that add an unpredictable field of their own; Mansfield and Chapman’s cure, in 1986, was active screening, a second winding that cancels the coil’s field outside it. And the coil is an inductor, so reversing a hundred amps in tens of microseconds takes a large voltage.

In 1991 few groups had such hardware. MGH’s scanner carried a resonant system from Advanced NMR Systems, in which a gradient coil rings against a capacitor and the amplifier supplies little more than the losses. Wong made the coil small instead. Days before that August meeting, over a weekend, he had built a head gradient coil from sewer pipe, wire and epoxy, meant for perfusion imaging; its low inductance let the scanner’s standard 100 A amplifiers reach about 20 mT/m in 50 µs. Minnesota had no echo-planar imaging and relied on field strength, since the susceptibility effect is stronger at high field; at 4 T it showed with conventional gradient-echo (FLASH) imaging. EPI became the workhorse anyway, and the major vendors put it on clinical scanners around 1996.

The last limit is the subject. A gradient switched fast enough induces enough electric field to fire peripheral nerves — magnetic stimulation by accident, the physics of TMS — and in 1993 Mansfield and Harvey studied the threshold with a small magnetic stimulator. Mansfield shared the 2003 Nobel Prize in Physiology or Medicine with Paul Lauterbur, who had shown in 1973 that field gradients could encode an image.

Why it took over

It requires no surgery, no radiation, and no tracer, so it can be used on healthy volunteers repeatedly. Combined with millimetre spatial resolution across the entire brain, that made human cognitive neuroscience into an experimental discipline rather than a mostly lesion-based one.

PET had been mapping activation in healthy volunteers since the mid-1980s, imaging blood flow after injections of short-lived tracers such as oxygen-15. Kwong’s flickering goggles had come, by way of Belliveau, from Peter Fox, who had used them for PET.

What came out of 1991 was a recipe as much as a result. Alternate a task with rest in blocks of tens of seconds, image continuously, and mark every voxel whose signal follows the alternation: Kwong subtracted rest images from task images, and Bandettini correlated each voxel’s time course with the task’s on–off boxcar. Change the task and the recipe maps what it engages; stack the slices, a volume every half-second to three seconds, and the map covers the brain. That is what the scanner bought neuroscience — whole-brain maps of what activates during a task, in healthy people, as often as an experiment needs. PET still dominated the first Human Brain Mapping meeting, in Paris in 1995; within a few years fMRI had taken the lead.

What the measurement actually is

Worth being precise about, because the gap between signal and inference is unusually wide here.

BOLD is a haemodynamic signal, not an electrical one. Neural activity increases local metabolic demand, blood flow over-compensates, oxygenation rises, and the MR signal changes — several steps of physiology between spiking and measurement. The response is slow, peaking some seconds after the activity that caused it and smearing anything faster, which caps effective time resolution at roughly the second scale no matter how quickly you sample.

The over-compensation is the whole effect. With PET in 1986, Fox and Marcus Raichle measured local blood flow rising 29% during somatosensory stimulation while oxygen consumption rose 5%. The surplus washes deoxyhaemoglobin out of the capillaries and veins, and the gradient-echo signal rises. If flow matched demand there would be nothing to see; why it does not has been argued over for decades.

And the activity the blood answers to is not quite spiking. In 2001 Nikos Logothetis and colleagues imaged the primary visual cortex of anaesthetised monkeys while recording from electrodes inside it. The local field potential — the slower part of the electrode signal, dominated by synaptic activity in the population around the tip — changed most, predicted the BOLD response better than multi-unit spiking did, and at sites whose responses were transient was the only signal that correlated with it significantly. Their reading was that BOLD reflects an area’s input and local processing rather than its spiking output, and the result held in alert monkeys in 2008.

Nor is the signal exactly where the activity is. In gradient-echo images the strongest signals come from veins on the cortical surface, which are unevenly spread and carry the change downstream, so a larger response need not mean more activity, nor an earlier one an earlier onset.

The change is small, and most of the noise is not the scanner’s. A 1.8% rise has to be picked out of frame-to-frame fluctuation of nearly 1%: the Wisconsin group were told in 1991 to expect a temporal signal-to-noise ratio of about 120, a figure Bandettini noted twenty years later had barely moved, because the fluctuation is physiological. Better electronics do not lower that noise floor; repetition does.

It is also relative. There is no absolute unit of BOLD, so results are contrasts between conditions, and the choice of baseline is part of the result.

This is the exact complement of EEG: where EEG measures the electrical signal directly but cannot localise it, fMRI localises well but measures something several causal steps removed. Neither limitation is an engineering deficiency to be fixed; both are properties of what is being sensed.