Functional magnetic resonance imaging, or fMRI, is a technology that measures brain activity by tracking changes in blood flow. This activity is captured as the blood-oxygen-level-dependent (BOLD) signal. Before this complex data can be interpreted, it must undergo a series of steps known as preprocessing. The purpose of preprocessing is to clean and standardize this raw data, removing interfering signals and artifacts to ensure subsequent statistical analyses are accurate and dependable.
Sources of Noise in fMRI Signals
The BOLD signal, which indicates neuronal activity, is subtle and can be easily obscured by various sources of noise. Non-physiological noise originates from the scanner hardware, such as thermal noise from electronic components that introduces random fluctuations. Another common issue is a slow, gradual drift in the signal’s intensity over the course of the scanning session, which is unrelated to brain activity.
Physiological noise stems from the subject’s own bodily functions. Rhythmic processes like heartbeat and breathing create cyclical patterns in the fMRI data that can be mistaken for brain signals. Chest movement during respiration and arterial pulsation with each heartbeat cause slight shifts in the brain’s position and the magnetic field, generating artifacts.
The most significant source of noise often comes from head motion. Even very small movements can introduce substantial errors into the fMRI signal. When a person moves their head, the scanner records data from fixed points in space, but the underlying brain tissue shifts. This causes large signal changes in a given voxel (a three-dimensional pixel) that can overwhelm the BOLD signal, creating false patterns of activation or masking true ones.
Correcting for Timing and Motion Artifacts
To address the issue of data being collected slice by slice, a process called slice-timing correction is applied. An fMRI scanner acquires a full brain volume by imaging it in a sequence of two-dimensional slices, meaning different slices are captured at slightly different moments. Slice-timing correction adjusts the signal in each slice so it appears as if all slices were acquired at the same time point, which is important for accurately modeling brain responses.
The next step addresses head movements that occur between the acquisition of each brain volume. This procedure, known as motion correction or realignment, involves aligning every volume in the fMRI time series to a single reference volume. Algorithms estimate the amount of translation and rotation that occurred between each scan and then apply transformations to correct for this movement. The process can be compared to stacking a pile of photographs and carefully aligning them so the final image is sharp.
Properly correcting these timing and motion artifacts is necessary for the analysis’s integrity. Head movements in particular can create signal changes that correlate with a task, leading to false conclusions about brain activation if not addressed.
Aligning Brains to a Standard Space
After correcting for initial artifacts, the fMRI data must be aligned to allow for meaningful comparisons. The first step is co-registration, which aligns the lower-resolution functional images containing the BOLD signal to a high-resolution structural image of the same individual’s brain. The structural scan provides detailed anatomy, allowing researchers to precisely locate brain activity within that person’s brain.
The next procedure is normalization. Since every person’s brain is different in size and shape, their data must be transformed to fit a common template to enable group-level analysis. Normalization is the process of warping each individual’s brain data to match a standardized brain atlas, such as the Montreal Neurological Institute (MNI) template.
This standardization allows researchers to average brain activity across many participants and report findings using a consistent coordinate system. Much like a globe provides a standard reference for Earth, a standard brain space allows for comparing activation locations across studies. Without it, comparing activation between individuals would be impossible.
Signal Enhancement and Filtering
The final preprocessing stage enhances the fMRI signal’s quality. One common technique is spatial smoothing, which applies a filter that averages the signal of each voxel with its immediate neighbors. This process slightly blurs the image but serves two main purposes. First, it increases the signal-to-noise ratio by averaging out random noise, and second, it helps the data better conform to the assumptions of later statistical models.
Another important step is temporal filtering, which removes unwanted frequencies from the signal’s time-course. For example, a high-pass filter removes low-frequency noise like scanner drift while allowing the higher frequencies associated with the BOLD response to pass through.
Once these steps are complete, the fully preprocessed data is ready for statistical analysis. Researchers can then model the data to identify brain activation during a task or explore connectivity between brain regions.