BME 499.098/Biostat 642:
Introduction to Functional MRI

Final Project Descriptions



Summary: The class has been divided up into eight groups, and we have have eight subjects' data. Each group will work on one of the projects defined below and will present the results on the last day of class.

Project presentation: Presentations will last 15 minutes with 5 minutes for questions. The group should take time to explain the nature of their project and the motivation for studying the question at hand. The results should be clearly presented, preferably with projection from a PC.

Due Date: Friday 8/19 in class.

Project I: Increasing Sensitivity with a Temporal Derivative Covariate

Overview: Including a temporal derivative of a predictor allows for some uncertainty in the temporal delay of the modeled events. In this project you will evaluate the impact of fitting of temporal derivative in one or more single subject analyses.

Points to Address:

  • Is there statistical evidence that the temporal derivative and or dispersion predictors are useful?
    Use a summary of F-statistic images decide
  • Does the temporal derivative subjectively improve the fit?
    Plot different voxels of interest
  • Does there appear to be a systematic temporal delay across voxels? If there is any such bias, is it similar across subjects?
    Look at the images of temporal derivative coefficients; create histograms. To create a histogram download histvol.m and save it into your spm2_umlocal directory.

    • Project II: Using a HRF basis to account heterogenity in HRF

    Overview: As discussed in class, the canonical HRF is the most powerful model only if it is accurate, if each subject's response has that shape. In this project you will compare HRF basis to see if a more flexible shape allows for improved sensitivity. You'll compare
    1. Canonical HRF
      (1 basis element)
    2. Gamma basis
      Window length 32s; Order 3 (3 basis elements)
    3. Fourier set (Hanning)
      Window length 32s; Order 3 (7 basis elements). (A 'Hanning window' ensures that the HRF starts at zero and ends, after 32s, at zero; there are 7 basis elements because there is 1 constant, 3 sines and 3 cosines.)
    4. Finite Impulse Response (FIR)
      Window length 32s; Order 16 (16 basis elements)
    In this project you will evaluate one or more single subject analyses.

    Recall that when you have a basis of HRF's, you need to use a F-test to test if there is experimental variability. See Tom for tips on constructing F-tests with lots of predictors.

    Points to Address:

  • Using F images, does the pattern of activation appear different between the 4 modeling approaches? If it differs, what is a possible explanation? (For comparability, use the same uncorrected F threshold, say P<0.01, for all images).
  • For the main region of activation (FFA), plot the same voxel in each of the 4 models. is the fit appreciably different in each?
  • For each one of the 3 basis HRF models (Gamma, Fourier & FIR) find an interesting significant voxel not seen on the canonical HRF results and plot it (make sure that you haven't found a bizzare edge or vasculature voxel). Are there important/intriguing results missed by the caonical HRF?
    • Plotting hints: Here are two ideas for plots: To show the fit over the whole time series, do
        Plot -> Fitted responses -> adjusted response
    To show the fit in a peristimulus time plot, do
        Plot -> Event-related responses -> 'fitted response and adjusted data'
    In all cases, when asked 'Which effect' always select 'Effects of Interest'.

    Project III: Relative Sensitivity of Levels of Inference & FDR

    Overview: SPM's tabular output provides corrected p-values for cluster- and voxel-level inference. In general, cluster-level P-values are the most sensitive and least specific, while voxel-level is the more sensitive and the most specific. This assessment, however, is based on general assumptions on the nature of the activations. In this project you will describe the relative sensitivity of these two types of inferences with yet a fourth, FDR voxel-level inferences.

    Points to Address:

  • On the class data, what is the relative ordering of the sensitivity of the three methods.
    Compare minimum corrected p-value, corrected intensity threshold, number of significant clusters, number of significant voxels.
  • Does the relative sensitivity (the order) of the different methods depend on smoothness?
    Try at least 3 different smoothnesses, e.g. 4, 8 and 16 FWHM.
    • Resources: Tool to view how FDR threshold is determined, use the FDRill function (a function in the spm2_umlocal directory). For each point above, evaluate as many subjects as possible.

    Project IV: Increasing sensitivity and specificity with motion parameters covariates

    Overview: Subject motion is the largest source of nuisance variability in fMRI. Even after motion correction, there can be variability that is explained by subject motion. In this project you will examine the impact of including motion parameters in an analysis.

    Points to Address:

  • Do the motion parameters explain a significant amount of variance?
    Use a summary of an F-statistic image.
  • Does adding the motion parameters increase or decrease significance? Give an explanation why either case could occur?
  • Do the motion parameters appear to be related to the experimental paradigm
  • Is there any correspondence between overall movement and data quality? For example, does the subject with the largest movements have the worst results?
  • If there is time, consider the following two points
  • Some authors argue that adding the squared (X.^2) motion parameters in addition helps.
    Use a summary of an F-statistic image to assess if adding squared motion parameters helps.
  • Other authors argue that adding that changes in motion are more important than the absolute motion. Instead of the raw parameters, try just including the temporally differenced motions (i.e. [0 0 0 0 0 0; diff(X)])
    Use a summary of an F-statistic image to assess if adding differenced motion parameters helps.
    • For each point, evaluate as many subjects as possible.

    Resources: The motion parameters can be found in the ra_img directory, in the file realign.dat. You can load a datafile such as this directly into an SPM prompt with spm_load(spm_get); this will bring up a file selection dialog, from which you can select the realign.dat file.

    Project V: To globally scale or not to globally scale

    Overview: A source of debate in the analysis community is whether fMRI data should be globally scaled or not. The "global" is the average intensity of all intracerebral voxels; since there is one global value for each scan, we have a "global" time series. "Global scaling" then is scaling each volume by its global value, such that all volumes have the same global average after scaling.

    The argument for scaling is as follows: Since we are interested in local changes, we would like to discount global changes; e.g. is it very interesting that voxel X increased by 5% if the entire brain also increased by 5%?

    The argument against scaling goes: Since each global value is just the average of all voxels, the global time series contains information about the paradigm, and by normalizing by the global we're attenuating our signal of interest. Worse, you can induce artifactual decreases. Further, the global signal that we're worried about is typically looks like drift, and hence any possible global-type signal will be removed by the drift basis.

    In Lab 1 & 2 you used scaling (remove Global effects: scale). Re-analyze your group's data without scaling, and re-analyze as many other group's data as possible (see if you can get the 'scaled' results from the respective groups).

    Points to Address:

  • Which analysis yields greater significance, with or with-out global scaling?
  • Subjectively, which analysis appears to better reflect the expected pattern of activation.
  • Using 'Check Reg', look at the pair unthresholded T images, one from a 'scaled' analysis, one from the 'unscaled' analysis. Is there evidence of artifactual de-activations in the globally scaled analysis?
  • A different approach to global normalization is to instead regress out the influence of the global time series, instead of scaling by it. Extract the global into a variable; go into the directory of an analysis which used global scaling. Type load SPM global = SPM.xGX.rg save global.dat global -ascii Now global.dat has the globals for your subject. Run a new analysis where you specify Other regressors (only one, actually). When asked for the regressor, you can type spm_load(spm_get) and go select the global.dat file.

    How are the global-regression results different from the global scaling results? Or from using no global normalization?

    • For each point, evaluate as many subjects as possible.

    If you'd like to read more about the global, a good starting point is GK Aguirre, E Zarahn,M D'Esposito. (1998). The inferential impact of global signal covariates in functional neuroimaging analysis. NeuroImage, 8, 302-306. Another article which has more current references is: Laurienti, Deactivations, Global Signal, and the Default Mode of Brain Function, J. Cogn. Neurosci..2004; 16: 1481-1483.

    Resources: The motion parameters can be found in the ra_img directory, in the file realign.dat. You can load a datafile such as this directly into an SPM prompt with spm_load(spm_get); this will bring up a file selection dialog, from which you can select the realign.dat file.

    Project VI: Impact of UM vs. SPM Preprocessing

    Overview: The UM fMRI Lab supplies investigators with preprocessed data, that is data that has had slice time and motion correction applied. These two corrections, however, are implemented with tools created outside of SPM*. In this project you will compare results created with data preprocessed with the standard UM tools with that preprocessed with SPM99's equivalent tools.

    * The UM slice time correction uses a locally windowed filter, while the SPM slice time correction uses a global filter (IFT(phaseshift(FT(X)))); the global filer may introduce artifacts. UM realignment (MCFLIRT) and SPM realignment both minimize the squared differences between the reference and transformed image, but MCFLIRT uses a more sophisticated optimization method and carefully deals with edge effects at the top and bottom of the brain.

    Points to Address:

  • Does the residual variance differ between the two processing paths?
    Compute ratios of ResMS images with ImCalc.
  • Do the activation results differ between the two methods? If so, how, and by how much?
  • If the results differ appreciably, find a particular voxel where the difference is striking. Plot the data for each one and try to suggest why they might differ.
    • For each point, evaluate as many subjects as possible.

    Project VII: Increasing sensitivity with a Gray Matter Mask

    Overview: Each statistic image of the brain comprises as many as 100,000 hypothesis tests. The corrected thresholds which control for the multiple comparisons problem must (naturally) get more stringent as the number of voxels increase. One suggestion to allow for less string thresholds is to eliminate voxels where activations are not expected, in particular in white matter regions. In this project you will restrict your inferences to a gray matter mask and describe it's impact on your results.

    Points to Address:

  • How do the corrected thresholds, voxel counts and RESEL counts change after application of the GM mask? Compare the "Euler Characteristic Density" with and without thresholding
    The EC density reflects the "crinkliness" of the search volume; the first 3 elements are small when the search volume is smooth. The EC densities are saved in the variable R in SPM.mat. Ask for help interpreting R
  • How do the activation profiles change with the mask; are there regions that disappear? Are there significant regions that appear?
    • For each point, evaluate as many subjects as possible.

    Resources: Gray Matter mask image: avg152T1_GM.hdr avg152T1_GM.img

    Project VIII: SnPM vs SPM

    Overview: The corrected p-values and thresholds in SPM are results from Random Field Theory (RFT). RFT depends on many assumptions on the data, most of them uncheckable. SnPM provides a nonparametric and data-driven method to obtain corrected thresholds. In this project you will apply SnPM to the group data and compare it to the results obtained with SPM.

    Points to Address:

  • Compare the intensity and cluster size thresholds of SPM and SnPM.
    Make sure the statistic values agree; the only thing different about SnPM is how corrected p-values and thresholds are obtained.
  • By smoothing the variance image the precision of local variance estimates can be increased by "borrowing strength" from neighbors. Try 10 mm FWHM variance smoothing; how does this change the results?

    • Project IX: SPM vs FSL

    Overview: FSL is another software package used to analyze fMRI data. While the methods are often similar, there are two key differences between FSL & and SPM's methods: At the "first" or single-subject level, FSL estimates temporal autocorrelation locally with a regularized ACF (autocorrelation function); SPM estimates a global autocorrelation model that approximates an AR1, rho=0.2 model. Second, FSL's group-analysis method uses information about within-subject variance to improve the final inference (in essence, "bad", high-variance subjects are down-weighted, and "good" low-variance subjects are up-weighted); SPM does not directly use the intrasubject variances, and hence cannot adjust individual subject's weights.

    This project is suitable for an advanced group, who has already used FSL or who is ready to quickly learn this different package.

    Points to Address:

  • Analyze the class' data as a group. This will require analyzing each subject individually, and then combining them in a 2nd level model.
  • For one subject, compare the intrasubject results for SPM and FSL. Are the statistic values very different? Are the FWE-thresholds very different (compare the RESEL counts and the FHWM smoothness)? If the thresholds are very different, use the same arbitrary threshold for both FSL's and SPM's statistic images.
    • Last modified: Wed Aug 17 15:36:36 EDT 2005