FEAT - Perfusion FMRI Analysis


Introduction

It is straighforward to analyse perfusion FMRI data (often also referred to as ASL or arterial spin labelling) with FEAT. The data needs special treatment because each voxel's timeseries alternates between "tag" and "control" conditions, with control timepoints having higher intensity. If there is BOLD signal present (for example because the echo time is not short) this can be modelled as the average of the tag and control. The perfusion (flow) signal can be modelled as the difference between the tag and control conditions; when there is activation, this difference increases.

There are currently two ways of analysing such data with FEAT - subtraction of tag from control (as part of pre-stats processing), or full perfusion signal modelling, where separate EVs model the BOLD signal, the (constant-height) tag-control difference and the modulation of this by the activation. The choice of which approach to take will depend on the data - subtraction can be better if there is certain structured noise in the data, whilst full modelling can give more accurate modelling, as the data does not have to be resampled, so no information is lost through interpolation.

With either approach, initial evidence suggests that it is worthwhile using highpass filtering (with cutoff set according to the same norms as BOLD FMRI analysis).


Full perfusion signal modelling

If you want to fully model the different aspects of the perfusion data, you will need to setup at least 3 EVs. See the example design matrix:

For an experiment with just a single experimental condition, these EVs are generally all that you need, and can be setup via the Model setup wizard. If you have more than one condition, then for each new condition you will need a new EV similar to EV2, modelling the BOLD response, and also a corresponding EV formed from interacting the BOLD response EV with EV1. Note that for each interaction EV, you need to set Make zero to Centre for EV1 only (i.e. leave the BOLD EV at Min). This makes the modulation envelope for the perfusion activation EV symmetric about zero, rather than only positive, and results in a slightly better model of the perfusion data.

The above assumes that the tag-control order of the data is such that the first timepoint is tag. If not, you should change the Phase for EV1 from 0 to the TR.

You should leave FILM whitening turned on.

Note that if you use this approach then featquery will not automatically output percent change values that relate to 100*(perfusion signal change)/(perfusion baseline). Instead you need to use featquery to output COPE1 (perfusion activation in the example above) and COPE5 (perfusion baseline in the example above), and then calculate 100*COPE1/COPE5.


Subtraction

Perfusion subtraction subtracts even from odd timepoints in order to convert tag-control alternating timepoints into a perfusion-only signal. If you are setting up a full perfusion model (see above) then you should not use this option.

To use the subtraction method, select your input data, and make sure that you set the TR correctly (i.e., how long each tag or control volume takes to acquire) in the Data part of the FEAT GUI. Then, in Pre-stats, setup the pre-processing as normal in FEAT, except that you should turn on Perfusion subtraction. You should not be using intensity normalisation so leave that turned off. You may wish to use Highpass filtering; this will remove low frequency noise still present after the subtraction processing. FEAT will automatically turn off FILM prewhitening if you use perfusion subtraction (this is set in the Stats part of the GUI).

By default it is assumed that your first timepoint is tag, the second is control, etc. If your data is the other way round, change the First timepoint is tag setting to First timepoint is control.

Next, setup the model as you normally would for an FMRI analysis, with an EV for each experimental condition. The subtraction causes a temporal shift of the sampled signal to half a TR earlier; hence you should ideally shift your model forwards in time by half a TR, for example by reducing custom timing onsets by half a TR or by increasing the model shape phase by half a TR.

To see exactly what the subtraction processing is doing, look at the $FSLDIR/bin/perfusion_subtract script. This splits the original timeseries into two, one containing all even timepoints and the other containing all the odds. It then creates two new timeseries from each, in one case shifting the data forwards by half an (original) TR and in the other shifting it back by half a TR. The appropriate combinations of these are then subtracted from each other, before re-combining into a processed dataset that has the same number of timepoints as the original data. Sinc (temporal) interpolation is used in order to minimised temporal blurring induced by the shifting. This approach has better accuracy (with respect to removing BOLD contamination and recovering perfusion signal) and efficiency (particularly with respect to not halving the number of timepoints) than simpler perfusion subtraction methods. However, the filter affects the autocorrelation structure in the timeseries, in a way that is not well-matched to the autocorrelation estimation in FILM. This is why FILM whitening is turned off, and instead, to correct for the reduced degrees of freedom in the filtered data, the varcope and degrees-of-freedom are automatically corrected after running FILM in OLS mode.

Note that if you use this approach then featquery will automatically output percent change values that correspond to 100*(perfusion signal change)/(perfusion baseline).