Non-parametric mapping tutorial

This tutorial guides you through a voxel-based morphometry. We will compare scans from patients with Temporal Lobe Epilepsy with controls who do not show epilepsy.

  1. You will want to install a copy of MRIcron – this also includes NPM
  2. This tutorial requires a sample dataset named npmdata.zip. This will be included on the DVD for members taking our course, for web users please contact Chris Rorden for this dataset.
  3. Our sample dataset includes images from 30 control participants (c1..c30.nii.gz) and 28 people with temporal lobe epilepsy (e1..e28.nii.gz). These scans are the smoothed normalized segmented gray matter maps (using a toolbox for SPM5 created by Christian Gaser). There is also a mask image (mask50.nii.gz), which is a binary file (black/white) showing brain regions that were more than 50% gray matter in the whole group (combining individuals with and without epilepsy). 
    graymatter

Compute results

Next, we we compute our statistical results. You could do this with SPM, but here we will use NPM which can conduct permutation thresholding and the non-parametric Brunner and Munzel test. To conduct these statistics, you need to download and install npm.exe – this only works on the Windows operating system.

  1. Launch NPM.exe.
  2. First go to the ‘Options’ menu and set the permutations to None. Permutation thresholding can be useful, but it will take at least 1000 times longer than a normal False-Discovery Rate corrected threshold, and is typically less sensitive. Therfore, for you first glance at your data, turn this feature off.
  3. Go to ’Option’ menu and click ‘Tests’ – make sure the ‘Brunner Munzel’ is checked and the t-test and Welch test are unchecked. Our data is normally distributed, so the t-test and BM test should give similar results.
  4. Click the File/ContinuousImagesBinaryGroups(VBM) command you will be asked to select three groups of images:
    • “Select Brain Mask” – select mask50.nii.gz
    • “Select Positive Group (Z scores positive if this group is brighter)” c1..c30.nii.gz – use the shift or control keys to select multiple images.
    • “Select Negative Group (Z scores negative if this group is brighter)” e1..e28.nii.gz – use the shift or control keys to select multiple images.
  5. You will be asked to name your output files – lets call them “conpos_epineg”.
  6. NPM will now compute the requested tests. It will create overlap images of all your patients (e.g. conpos_epinegMn.nii.gz) and a statistical map (conpos_epinegttest.nii.gz for the t-test, conpos_epinegBM.nii.gz for the BM test).
    NPM

    7. Viewing results

    We can open up the statistical maps generated and place them on top of an anatomical scan. If your lesion maps were aligned to stereotaxic MNI space, you can open them on top of one of the standard templates (File/OpenTemplates/ch2). Here is a quick guide:

  • Launch MRIcron and choose File/OpenTemplates/ch2bet as our background image
  • Choose Overlay/Add and choose the statistical map created in the previous step (e.g. C:\dataset\binL.nii.gz).
  • When MRIcron detects a statistical map, it calculates the p-values for each test in order to determine the false discover rate (FDR) threshold – e.g. how much robust signal is present in your data. MRIcron displays a histogram of the Z-scores. Note in the example below, most of the data has positive Z scores suggesting a robust signal (if our data was merely noise, we should see a bell-shaped distribution with a mean of 0, instead the mean Z score is around 2).
    Histo
  • Next, MRIcron displays the overlay. Note at the bottom of the screen the software reports the critical values: the p05/p01 values correspond to uncorrected p<0.05 and p<0.01 values that are very liberal (these tests will make many false alarms, e.g. here we have conducted 1100647 tests, so p05 should result in many false positives). The fwe05 and fwe01 values correspond to the Bonferroni-corrected values: this test is very conservative, and you will often fail to detect real effects. The FDR05 and FDR01 results reflect the False Discover Rate – e.g. a FDR05 should show around 20 real activations for every false positive. Note that when there is very little or no signal, FDR is as conservative as Bonferroni, but it is adaptive to the actual signal in your dataset. Note that you can select the image thresholding and cutoff for your overlay, as described in the MRIcron manual. Note that by default, my software loads statistical maps with thresholds from FDR05 to FDR01, unless there is insufficient signal, in which case it uses the uncorrected 0.05…0.01 values. Also note that the current overlay is set to appear in a monochromatic red color scheme.
    mricron
  • We can repeat steps 2..4 to load multiple overlays, to compare different statistical tests. For example, clicking on the ‘tutorialfmri.bat’ icon will launch MRIcron and load to overlapping regions of interest. By using the Overlay/TransparencyOnOtherOverlays command we can view both of these overlays simultaneously.