Mapping brain activity with fMRI

Functional magnetic resonance imaging (fMRI) does not record the activity of neurons directly. When a brain region becomes more active, the local blood flow rises by more than the extra oxygen the tissue consumes, so the blood leaving the region carries less deoxyhaemoglobin. Deoxyhaemoglobin is paramagnetic and slightly disturbs the magnetic field, which weakens the MR signal; when there is less of it, the signal rises by a few per cent at most. This is the blood-oxygen-level-dependent (BOLD) signal.

The BOLD response is slow. After a brief burst of activity it rises over several seconds, peaks approximately five to six seconds later and falls back, often with a small undershoot, over about twenty seconds. This curve is the haemodynamic response function (HRF). Because responses to successive events overlap and add up, the timing of an experiment decides how well the response can be measured. Analyses usually predict the signal by convolving the timing of the task with the HRF and fit a general linear model to every voxel (Friston et al., 1998); randomising the intervals between events greatly improves event-related designs with short intervals (Dale, 1999).

A brain image contains tens of thousands of voxels, and a test is made at each of them. At an uncorrected threshold of p < 0.05, about one voxel in twenty would pass by chance alone, so whole-brain analyses correct for multiple comparisons, controlling either the family-wise error rate or the false discovery rate. Conclusions about people in general come from a group analysis in which each participant contributes one estimate.

The session maps the hand area of the motor cortex while a simulated volunteer taps the fingers of the right hand. The task design, the haemodynamic response, the noise and drift, the general linear model and the power calculation follow the GloNeuro fMRI design simulator. The brain slice, head motion and its artefacts, spatial smoothing and the group of participants are simplified illustrations, with approximate values. Unlike in a real scan, the true location and size of the activity are known, and the final station shows how each decision shaped the map.

Learning objectives

• Explain why the magnet of an MRI scanner is always on, why ferromagnetic objects are dangerous, and why implants are checked by trained MR staff before anyone is scanned
• Position a participant with a head coil, padding, hearing protection, an alarm and a response box, and explain how each affects the data or the participant's safety
• Describe how the BOLD signal arises and why the haemodynamic response peaks approximately five to six seconds after brief activity
• Choose a block or event-related design, predict its signal by convolution with the HRF, and compare designs by their efficiency
• Set the repetition time and run length, and relate them to the number of volumes
• Correct head motion, choose spatial smoothing, fit a general linear model at every voxel, define a contrast and threshold the t map with uncorrected, family-wise error and false discovery rate thresholds
• Run a group analysis with the participant as the unit of analysis and interpret the power of the study

Ethics and safety

Research with people requires approval by an ethics committee (in India, an Institutional Ethics Committee following the ICMR National Ethical Guidelines of 2017; elsewhere, a research ethics committee or, in the United States, an institutional review board) before recruitment, in line with the Declaration of Helsinki, and written informed consent after the participant has read an information sheet and had the chance to ask questions; participants may withdraw at any time without giving a reason. MRI involves no ionising radiation, but the static magnetic field of the scanner is always on, so every person entering the scanner room is screened for implants, metal fragments and ferromagnetic objects, and the screening is checked by trained MR staff (a radiographer or MR safety officer) who decide whether scanning is safe. The scanner is loud, so hearing protection is required; participants hold an alarm they can squeeze at any time; and some people find the bore too confined. Research scans are not diagnostic, and the information sheet explains what happens if an unexpected finding is seen. Data are stored under a participant code, with the consent form kept separately.

Key terms

• MRI safety screening: A questionnaire and interview, checked by trained MR staff, that asks about implants, metal fragments, previous surgery, pregnancy and other conditions before anyone enters the scanner room.
• Ferromagnetic: Strongly attracted by a magnet, like iron and most steels. Ferromagnetic objects can be pulled into the scanner at high speed.
• Head coil: The receiver that surrounds the head and picks up the MR signal; it also holds the foam padding and the mirror through which the participant sees the screen.
• BOLD signal: Blood-oxygen-level-dependent signal: the small rise in the MR signal that follows neural activity, because blood flow increases and deoxyhaemoglobin falls.
• HRF: Haemodynamic response function: the time course of the BOLD signal after a brief event, peaking approximately 5 to 6 s later.
• Block and event-related designs: In a block design the task is performed continuously for tens of seconds, alternating with rest; in an event-related design brief trials are presented separately, at fixed or jittered intervals.
• Convolution: Adding up the HRF once for every moment of stimulation, shifted to that moment; it turns the task timing into the predicted BOLD signal.
• Design efficiency: How precisely a design lets the response be estimated; higher is better.
• TR: Repetition time: the time to acquire one whole-brain volume. A run of length T contains T ÷ TR volumes.
• Voxel: The smallest box of volume in the image, here 3 mm on each side.
• Realignment: Motion correction: each volume is moved back into line with a reference volume, and the estimated movements are kept as motion parameters.
• Spatial smoothing: Averaging each voxel with its neighbours using a Gaussian kernel of a given width (FWHM, full width at half maximum).
• GLM, β and contrast: The general linear model explains each voxel's signal as a weighted sum of regressors plus noise; β is the weight of a regressor, and a contrast is the combination of βs that is tested.
• FWE and FDR: Family-wise error control limits the chance of any false positive among all voxels; false discovery rate control limits the expected proportion of false positives among the voxels declared active.