Basic fMRI analysis
In this exercise you should do a simple statistical analysis on fMRI data using the general linear model, as implemented in matlab (glmfit). Write a report describing and showing your findings.
The data is from a single subject performing a simple auditory experiment. The auditory stimulation was words presented in a block design (to both ears) at a rate of 60 words per minute. 96 acquisitions (or scans) were made with a TR (repetition time) of 7s. There were two experimental conditions: auditory stimulation and rest. The conditions were presented in blocks of 42 seconds (6 scans), giving a total of 16 blocks. The condition for successive blocks alternated between rest and auditory stimulation, starting with rest. The first 12 scans (one task and one rest block) were discarded, leaving 84 acquisitions.
The data has been preprocessed: The functional images were realigned (motion corrected), and coregistered to the anatomical image. The images have been normalized to the MNI template brain. Finally, the functional data was smoothed with a 6-mm FWHM Gaussian kernel.
For the purpose of your analysis (and to reduce the amount of data posted on the Moodle page) I have extracted the preprocessed data from three different regions of interest (ROIs): the left auditory cortex, the right auditory cortex, and an unrelated region in the left posterior cortex (~150 voxels in each). Your task will be to:
1) Create and show a 'boxcar' model of the experiment based on the description above.
2) Take the shape of the BOLD response into account and create a better model for your experiment. Use the shape of the hemodynamic response function (sampled at 1s) from the mat-file containing the hrf response. Plot the result.
3) Perform a statistical analysis on each voxel in the three different ROIs using the general linear model. Can you find statistically significant results? A p-value of 0.001 (for the gain) can be considered significant for the uncorrected data.
4) Correct for multiple comparisons. You may use Bonferroni correction. Do your results survive correction? A corrected p-value of 0.05 can be used.
5) How well does your model fit the data in i) a voxel with a significant result and in ii) a voxel where you did not find a significant result? Plot the responses, the fitted models, and the residuals.
6) How could you further improve your model? Describe (in the text) at least one nuisance factor that you could include in your design matrix to get a better fit.
Useful matlab commands: conv, glmfit, glmval