3 Data Preprocessing

  • Motion correction
  • Slice time correction
  • Spatial filtering
  • Intensity normalization
  • Temporal filtering

List 1:Commonly used data pre-processing steps in functional MRI analysis.

The data pre-processing steps listed above (List 1) are routinely carried out prior to statistical analysis, yet it is often the personal preference of a particular researcher (or the defaults of the software package that he/she uses) that determines which steps are eventually carried out and in what order. How pre-processing is carried can also depend on characteristics of the data acquisition. For example, fMRI data can be acquired with a motion correction option that is available on the MR scanner. If motion corrected data is acquired, it not necessary (but still optional) to carry out motion correction as a part of the data preparation steps for statistical analysis.

It is a safe bet to adhere to the default settings of common statistical analysis packages; they represent peer-reviewed accepted pre-processing implementations. However, to obtain valid and optimal results, it is important to know about the details of your data, for example, the repetition time (TR) the slice acquisition order, and if motion correction has been applied directly during scanning.

Motion correction

For a valid analysis of functional MRI data, it is assumed that the variation of intensity in the same voxel between repeated measurements (volumes) of a time-series is primarily due to changes in cerebral physiology. This means that the identical set of voxels time-courses should be sampled during the entire time-course of the experiment. For a given voxel, head motion results in shifting the signal time-courses of neighboring voxels in and out of the studied voxel, rendering the assumption that a single time course for this voxel is studied false. The problem of motion is most apparent in voxels that are located on edges between high and low intensity tissue, such as around the cerebrospinal fluid filled ventricles in the brain. To overcome the problem of motion, methods to correct motion artifacts have been developed and are now widely used for MRI analysis.

There are different sources of motion in an MRI experiment. The easiest motion source to imagine is the movement of the head during the time-course of the experiment. People may move because they feel uneasy in the headcoil, but also because they are engaged by the experimental task that they perform. Apart from these factors, motion is caused by respiration, but also pulsation of the blood stream causes shape changes (motion) in the brain. Together these factors provide a complex set of parameters that can be taken into account when trying to correct for motion related errors.

Most motion correction methods assume that the head is an object that doesn't change shape in the MRI scanner. This means that such methods only correct for rotations and translations along the x, y and z axes that define a given voxel. This is referred to as a rigid body registration approach with 6 degrees of freedom (3 rotations and 3 translations). However, non-rigid shape changes in brain tissue do occur. For example due to pulsation of the blood stream or because motion occurs in the time that a signal from a slice is sampled, which results in an apparent shape change. These motion parameters are not accounted for by most current motion correction methods.

Motion correction is actively being researched. The links below will bring you to the abstract of several current journal articles in PubMed. Downloading the full article may require a subscription.
Barry RL, et al. 2009. Magn Reson Imaging, epub ahead of print
Costagli M, et al. 2009. NeuroImage, 45(3):749-57
Morgan VL, et al. 2007 Comput Med Imaging Graph, 31(6):436-46
Speck O, et al. 2006. MAGM, 19(2):55-61
Bannister PR, et al. 2006. Image and Vision Computing, 25(3):311-20
Bannister PR. 2003. PhD thesis on motion correction.

LinkOut: Further reading about motion correction.

Slice time correction

Scanning functional MRI volumes normally occurs in 2 dimensions (2D) meaning that only one slice is acquired at a time. In effect, there is a small time difference between acquisition of the first and the last slice in the volume. During this time lag, physiological differences occur, such as differences in the heart rate, respiration and as a result head motion and changes in the functional signal of interest. Without correction, the Blood Oxygen Level Depended (BOLD) response for a single event appears to start earlier in time for the last slice in the volume compared to the first slice in the same volume. Furthermore, the analysis of FMRI data assumes that information in one fMRI volume is acquired at the exact same moment, otherwise the statistical model which is used to describe the data will fit less optimally. To compensate for the difference in acquisition time between slices in a volume, the time-series of each voxel in a slice can be shifted slightly forward or backward in time. This is referred to as phase shifting of the time-series or simply slice-time correction. Although slice-time correction was considered an essential step in data-preprocessing, fMRI data analysis packages have switched off this option as a default. The SPM 8 manual says about slice-time correction: "Note that this option is likely to be removed in future. The authors of SPM do not generally suggest that this correction should be used, but the option is still retained for the few people who like to use it." The reason not to perform slice-time is subject motion. Motion causes the image to be distorted and an unpredictable phase-shift of the time-series per slice. This may be overcome by carrying out motion correction and slice-timing correction at the same time, but currently this is not implemented in most analysis packages.

Kiebel SJ, et al. 2007. Neuroimage, 34(4):1487-96

LinkOut: Further reading about slice time correction.

Spatial filtering

Spatial filtering involves blurring of the functional MRI data and its main goal is to remove noise while at the same time the signal of interest is retained. In general, high spatial frequencies in fMRI data are likely to represent noise components in the data, whereas low frequencies are changes produced by blood flow and hence are more likely to represent the signal of interest in Blood Oxygen Level Dependent (BOLD) imaging. Since smoothing retains the signal whereas noise is removed, one can say that the signal-to-noise ration increases through smoothing. Another reason to apply spatial smoothing is that the subsequent statistical analysis requires that data is normally distributed. Smoothing the data has a normalizing effect (4mm FWHM is considered adequate for statistical normalization; see below). Finally, individual differences in brain anatomy become less pronounced by smoothing and hence a cross-subject analysis is more valid.

Spatial filtering is carried out by convolving a 3D volume with a Gaussian kernel. A Gaussian kernel is a symmetric bell shaped curve. Convolution is a mathematical operation between the two signals (i.e. 3D volume and Gaussian kernel) that results in a single third signal (Figure X). Smoothing should not be carried out when small activations are expected, since activations smaller than the chosen filter size will not be detectable. The smoothing filter is defined by its Full Width Half Maximum (FWHM), referring to the width of the Gaussian curve at the half of its maximum. A FWHM kernel between 3 and 10 mm is commonly used for fMRI images.

Convolution of two square functions. Image Source: Wikipedia

Intensity normalization

Intensity normalization results in a change in brightness of all functional MRI images in a time-series, such that they all have the same mean intensity. Intensity normalization is common practice in Positron Emission Tomography (PET) analysis, since the signal changes greatly over time due to the decay of the radioactive signal that originates from the injected radioactive agent. In functional MRI, the signal varies much less and problems in the statistical analysis may arise when intensity normalization is applied. This can be understood by imagining a fMRI volume in which a cluster of voxels has a very strong response. Such a volume has a higher than normal average value and rescaling it will result in some of the less active voxels to show a decreased (negative) fMRI response. This appears as a deactivation to the stimulus in the statistical maps that are calculated later. When analyzing data with FSL, intensity normalization is turned off by default and discouraged. To overcome the problem of varying intensities between volumes, a grand mean scaling is often carried out. The scaling factor for grand mean scaling should be a value that is less sensitive to outliers than the mean intensity across the time-series. For example, the median intensity over all volumes in the time-series could be taken.

Murphy K, et al. 2009. Neuroimage, 44(3):893-905
Gavrilescu M, et al. 2002. Neuroimage, 17(2):532-42
Andersson JL. 1997. Neuroimage, 6(4):237-44

LinkOut: Further reading about intensity normalization.

Temporal filtering

Temporal filtering is carried out to remove unwanted temporal frequencies of a time series. Two types of filters have been commonly used, although only the first method continues to be used by most fMRI researches. This first method, a high pass filter, can be used to remove lower frequencies from the fMRI time-series. The filter allows high frequencies to pass through the filter, whereas the low frequencies are taken out. The slowly varying signals can be related to heartbeat, respiration or scanner noise. Removing these components will increase the signal-to-noise ratio. More advanced methods to remove cardiac and respiratory noise from fMRI time-sequence data are under development (see LinkOut of this section). Second, a low pass filter can be used to reduce high frequency noise from the time-series. However, the risk of losing the signal of interest is high and high pass filtering may introduce temporal autocorrelation. The problem of temporal autocorrelation is explained in the statistics paragraph. New methods for denoising fMRI time series data are actively researched (see LinkOut of this section).

Friston KJ. 2000. Neuroimage, 12(2):196-208
Deckers RH, et al. 2006. Neuroimage, 33(4):1072-81
Hagberg, GE, et al. 2008. Magn Reson Imaging, 26(7):1026-40
Stausberg S, et al. 2009. Phys Rev E Stat Nonlin Soft Matter Phys, 79(4)
Yan L, et al. 2009. Magn Reson Med, 61(4):819-27

LinkOut: Further reading about temporal filtering.

Continue Reading