Jessica Lammert & Kate Raymond (Fall 2019)

Goals

• To understand the differences between whole-brain voxelwise and ROI analysis approaches
• To learn how to set up different types of ROI analyses for analyzing group data
• To appreciate the advantages and disadvantages of functional and meta-analytic localizers
• To understand the problem of inter-subject variability and the usefulness of individual localizers

Relevant Lecture

Group Data

Accompanying Data

External Hard Drive (folder: 9223project_ROItutorial_JL+KR)

Background

One of the primary goals of fMRI is to determine which brain regions respond to certain stimuli or become active during a given task. So far in these tutorials we have used whole-brain voxelwise analysis to assess brain activity. However, many researchers opt to extract the BOLD signal from a specified brain region called the region-of-intest (ROI). In this tutorial we will examine the utility of the ROI approach for analyzing group data in contrast to the whole brain voxel-wise approach. In addition, we will cover three methods of defining ROIs:

1: Group neurosynth localizer

2: Group functional localizer

3: Individual functional localizers

To do this, we will be using the experimental data from 13 particiapnts. This data has already been pre-processed for motion correction as well as spatial and temporal smoothing. Here, we will compare the anatomical placement of different types of ROIs and run a general linear model (GLM) in each to compare the statistical output.

Overview: Voxel-Wise vs ROI Approach

We have already generated a Whole Brain Voxel-Wise RFX GLM for the experimental runs. Below is the anatomical avereged brain (13 participants) and the corresponding GLM overlayed with the contrast we are interested in: Faces > Hands. Recall that this map is created by computing an ANOVA in each voxel of the brain to test whether there is a significant main effect of Faces.

Question 1: Looking at the heat map above (where we simply used a statistical threshold of p=0.05), in what brain regions do you see significant activation for Faces > Hands?

Since we have run this GLM in over 300,000 voxels in each participant, we have to correct for multiple comparisons. Recall that there are various ways to apply this correction in BrainVoyager. In the following steps you will import the same GLM as above but where the False Discovery Rate (FDR) method has been applied to correct for multiple comparisons.

1) Open BrainVoyager, then click File Open and navigate to the Tutorial Data Folder, then select AverageAnat_P2_to_P14.vmr and click Open.

2) Select the Analysis menu and click Overlay General Linear Model.

3) A window titled "Overlay GLM Contrasts" will appear. Click Load.GLM... and select the voxelwise.glm file, then click Open. Now enter in the contrast we are interested in by selecting [+] for Face_Left, Face_Center, and Face_Right and [-] for Hand_Left, Hand_Center, and Hand_Right, then Click OK.

Question 2: How does the brain activity after adjusting the statistical threshold compare to brain activity before correcting for multiple comparisons?

In the voxel-wise approach we tested the hypothesis that Faces > Hands in every voxel of the brain. In contrast, the region-of-interest (ROI) approach tests the hypothesis that Faces > Hands in a pre-defined brain region.

Question 3: Statistically, what is the main benefit of applying an ROI approach compared to a voxelwise approach? How might your study design influence your decision to use the ROI approach over the voxelwise approach and vice versa?

Group Neurosynth Localizer

The easiest way to select a ROI is to choose a stereotaxic coordinate based on prior studies or brain atlases and draw a sphere of a given size around that point. ROIs based on single studies can be more sensitive to noise, so a better way to define an ROI is from meta-analyses of the domain or task of interest. Neurosynth is a database of fMRI studies that combines thousands of published results to create activation maps. You can search Neurosynth for a region or function of interest to see the results of all studies in the database under that key term.

1) Go to https://neurosynth.org/

2) Click Locations, then Associations, type any brain region or function of interest into the search bar to view the meta-analysis results.

In this tutorial we are interested in the contrast Faces>Hands and will using the right fusioform face area (FFA) as our ROI. We have already extracted a Neurosynth heatmap for the FFA based on a meta-analysis of 99 studies. In the following steps you will overlay that heatmap.

3) In BrainVoyager, under the Analysis tab, click Overlay Volume Maps and Load... the file neurosynthloc.vmp.

4) Looking in the right hemisphere, use the increase threshold button (toolbar on the left) to increase the statistical threshold until only peak voxel(s) activation remains.

Question 4: What is the approximate coordinate of the voxel(s) that survived increasing the threshold? Does this location make sense given your knowledge of the FFA?

We used this coordinate in a MATLAB script to create an ROI of 1416 voxels, which consists of the neighboring voxels (extending from the peak voxel in any direction) with the most activation.

Question 5: What might be a problem with selecting too many voxels for our ROI? How might the number of voxels change depending on the region you are interested in?

[Note: BrainVoyager refers to a Region-of-Interest (ROI) and Volume-of-Interest (VOIs) interchangably, they are the same thing.]

10) Select the VOI GLM tab at the top of the window, then click VOI GLM button.

11) In the new ROI GLM Specifications window, go to the RFX tab at the top of the window and ensure RFX GLM is checked.

12) Then click the Contrasts tab and set the contrast to [+] for Face_Left, Face_Center, and Face_Right and [-] for Hand_Left, Hand_Center, and Hand_Right, BrainVoyager should automatically apply this contrast for all of the participants.

13) Click Fit GLM, it may take a few seconds for the GLM to compute.

14) Observe the RFX GLM results. Pay particular attention to the contrast statistics in the ANOVA window (e.g. the p value). You can leave this table open or save it for later comparison.

Group Functional Localizer

We can follow a similar approach to define an ROI using functional data collected from our own participants. Recall that for all participants, we collected 2 runs of a localizer task where participants viewed Face, Hand, Body, and Scrambled stimuli and 6 runs of an experimental task where participants were presented with Faces and Hands in Left/Right/Center orientations. We created a group functional localizer ROI following the same method as the Neurosynth ROI above but, instead of meta-analytic data, the ROI was defined with a whole-brain voxelwise RFX GLM for the contrast Faces > Hands using the group localizer data.

Question 6: We created the ROI with the group localizer data and will run the ROI GLM with the group experimental data, what would be the problem with running the ROI GLM on the same data that we used to define the ROI?

1) Select Analysis, Region-Of-Interest Analysis, then click Add... (not Load...) and select the groupfunctlocffa.voi to add the functional localizer ROI to the list.

2) Click FFA Group Functional and Show VOIs to see this ROI on the brain, Shift-click to select the group functional and Neurosynth ROIs from the list and click Show VOIs to see multiple at the same time.

Question 7: Why might the location of the two ROIs differ?

3) Repeat steps 9-14 in the Group Neurosynth Localizer section to perform an RFX GLM in the group functional localizer ROI, make sure you select FFA Group Functional from the VOI list in Step 10.

Compare the results of the RFX GLM for the Group Functional ROI to the table produced by the RFX GLM in the Group Neurosynth ROI.

Question 8: Why might the statistical results (from the ANOVA tables) differ between the Neurosynth ROI and the Group Functional Localizer ROI? Which method appears to be more statistically powerful?

Individual Functional Localizers

Funcational localizers follow the assumption that the voxels defined as the ROI in the localizer scan are the same voxels activated in the experimental scan. However, inter-subject variability may compropmise this assumption. Now we will make a ROI for an individual subject (P02). This ROI is created similarly to the group data (using localizer runs to find the region that responds to Faces>Hands), but is tailored to the individual. So here we will first run a whole-brain voxelwise GLM with the localizer data for P02.

1) File, Open, then navigate to the P02 folder and select P02_Anat-S1_BRAIN_IIHC_MNI.vmr and click Open.

2) Select the Analysis menu and click General Linear Model: Multi Study, Multi Subject.

3) The corresponding window will appear. Click Load.GLM..., select the P02_Loc.mdm file, then click Open. The VTC files should now be imported, so hit GO to start the localizer GLM. This may take a few seconds to compute.

Now that we have run the GLM, we will overlay the contrast Faces > Hands.

4) Select the Analysis menu and click Overlay General Linear Model. Select [+] for Face, [-] for Hand, and [ ] for all other predictors. Then click OK to visualize the contrast.

For this tutorial, we have been using the right FFA as the ROI. In order to define this region of interest in subject P02, you can first find the peak voxel (the voxel with the most significant activation) in the right FFA like we did with neurosynth.

5) Navigate the crosshairs to the region broadly correspoding to the right FFA. Then, increase the threshold until you find the peak voxel.

Question 9: What are the X, Y, and Z coordinates for the peak voxel you chose? Are these the same coordinates defined by Neurosynth? Why or why not?

The next step is to define the voxels that constitute the ROI by determining the neighbouring voxels with the most significant activation around the peak voxel coordinate. We have already done this using a MATLAB script in the same way we defined the Neurosynth and Group Functional ROIs.

Now that we have defined the ROI for subject P02, you can import it using the same method as the Group Neurosynth ROI and Group Functional ROI.

6) Back on the AverageAnat_P2_to_P14.vmr tab (not the P02 tab), select the Analysis menu and click Region-Of-Interest Analysis . When the corresponding window pops up, select Add... and click indivfunctlocffap02.voi and then click Open.

7) To visualize the ROI for P02 click P02_FFAright and then Show VOIs .

Question 10: Compare the group functional localizer to the localizer from P02. How are they anatomically similar or different?

We repeated this process to create an ROI for each subject. Next, you will import these ROIs to visualize the inter-subject variability.

8) Go back to the Region-of-Interest Analysis window. This time, select Load... and click allvois.voi and then click Open.

9) To visualize all these ROIs at once, Shift-click to select them all and hit Show VOIs .

Question 11: Each subject's ROI was defined using the same method. So what might be causing the variability displayed here?

Sometimes we may want to quantify the spatial consistency among individual ROIs, we can do so with a Probability Map . We have already created a probability map for you to import into BrainVoyager.

10) Select Analysis, click Overlay Volume Maps , and then

Load... the allindividualsprobability.vmp file.

11) Navigate the crosshairs to the right FFA and observe the heatmap, you may have to Hide VOIs in the Volume-Of-Interest Analysis window to visualize it more clearly.

This map shows the extent to which all 13 individual ROIs overlap. The orange/red voxels represent greater overlap of individual ROIs or regions where activity is more consistent across participants. In contrast, the lighter yellow voxels represent lower overlap of individual ROIs or areas where there is more inter-subject variability.

In this heatmap, F = 1 indicates 100% overlap. So here, the voxel with the most overlap among the ROIs is 7/13 participants or F = 0.54.

Question 12: Where does the most overlap occur? What does the degree of inter-subject variability tell you about the Group Functional ROI?

This probability not only illustrates the inter-subject variability but also demonstrates another problem with using functional localizers to define an ROI. The FFA, like many regions in the brain, is functionally heterogenous. In other words, all of the voxels in the FFA do not necessarily respond to stimuli the same way, instead it contains smaller functional subdomains that may respond differently to different types of stimuli. Here we used the same contrast (Faces>Hands) in both defining the ROI (using localizer runs) and computing the GLM (using experimental runs). However, many experimenters utulize a different contrast in the experimental run than in the localizer run, which likely exacerbates the problem of functional heterogenity.

Now that we have created an ROI for each of the 13 participants using the localizer data, we can compute an RFX GLM for the experimental data using the individual subject ROIs.

12) Deselect the Probability volume map.

13) In the Volume-of-Interest Analysis window, select Options...

14) When the corresponding window appears, under the Access Options tab ensure that the following options are selected: Subj_VOI naming convention, Use subject's VOIs for time course access, and Separate subject predictors. If it isn't already loaded, hit Browse to select a design matrix file, and select Multi-Participant_Exp.mdm and click Open.

15) Then, under the VOI GLM tab click FFAright in the VOI list and then the VOI GLM button.

16) In the corresponding window "ROI GLM Specifications" click the RFX tab and make sure Separate subject predictors and RFX GLM are selected. Then, under the Contrasts tab make sure Face conditions are set to [+] and Hand conditions are set to [-], then click the FIT GLM button.

Now that we have computed a GLM in the Group Neurosynth, Group Functional, and Individual Functional ROIs, we can compare the statistical outputs. Look at the ANOVA tables for each GLM, pay particular attention to the contrast statistics, including the p-value and error term.

Question 13: How do the different methods for defining an ROI impact the statistical significance? Which method do you think is most statistically powerful?

Question 14: For each of the methods for defining an ROI covered in this tutorial, discuss a strength and weakness.