Max
Back to Psych 204 Projects 2009
Background
fMRI-Adaptation (fMRI-A) refers to the phenomenon that, as a region of cortex is activated by a stimulus repeatedly, it becomes less responsive overall to successive presentations of the stimulus. While the exact mechanisms of adaptation vary from region to region, fMRI-A in visual cortex has been studied in depth (see Grill-Spector et al., 2006, Weiner et al., 2010). In order study fMRI-A, stimuli must often be presented in rapid succession using event-related designs. These designs allow experimenters to study temporally shorter processes, but they come with several methodological complications, specifically in data analysis techniques.
The shorter inter-stimulus intervals involved in event-related designs lead to signal overlap across different trial types and may induce dependencies and nonlinearities between study events and/or conditions. Because of these issues, standard General Linear Model (GLM) analyses may be less precise and appropriate than usual. Choices as to how to analyze timecourse data in event-related designs can have important consequences for significance detection, data visualization, and relative activations to different types of stimuli. We hope to explore these differences in the current study.
Region of Interest (ROI) GLM analyses provide crucial information about regional selectivity and activation, yet they oversimplify data by averaging activation across an entire region down to a single timecourse. Multi-voxel patterns can be used to further investigate and characterize the distribution of activity within a region by cross-correlating activation on a voxel-by-voxel basis between trial types. In ROI analyses, different voxels may drive the same signal change within an ROI. Using MVPA, distributed patterns of activation may be correlated and one is able to test effects that a GLM would not be able to capture.
A second goal of this project was to familiarize the author with standard fMRI data analysis techniques.
Methods
Study Design
Faces, body parts, and other objects were shown to participants in the scanner for 1,000 ms per stimulus. Some stimuli were repeated (6 times) and others were presented only once. Repeated trials had an average of 20.44 trials between repeats. Trial types (repeated vs. non-repeated vs. blank vs. scrambled image) were counterbalanced and the repeated trials were counterbalanced within trial type. Subjects also performed a categorization task during the study, pressing a different button for each stimulus category.
Scan Properties
MR acquisition
Data were obtained on a 3-Tesla GE scanner. 12 slices were acquired at a resolution of 1.5 x 1.5 x 3.0 mm. TR was 1,000 ms and TE was 30 ms. Slices were acquired using a two-shot T2*-sensitive spiral acquisition sequence. The flip angle was 77 degrees and the field of view was 192 mm.
MR Analysis
The MR data was analyzed using mrVista software tools.
Pre-processing
Data were pre-processed (motion corrected, slice time corrected, etc.) by Weiner and colleagues.
HRFs vs. Deconvolution
We analyzed 2 face-selective ROIs and one body part-selective ROI. Standard GLM analyses were run with several different built-in HRF functions (Boynton, Dale & Buckner, and SPM) as well as with deconvolution and results were compared and contrasted.
Multi-Voxel Patterns
Multi-voxel patterns were examined for the same ROIs.
Results
HRFs vs. Deconvolution
Face-Selective Region using Boynton HRF
Face-Selective Region using Dale & Buckner HRF
Face-Selective Region using SPM Difference-of-Gammas HRF
Face-Selective Region Using Deconvolution
Multi-Voxel Patterns
Face-Selective Region #1
Face-Selective Region #2
Face-Selective Region #3
Conclusions
HRFs vs. Deconvolution
Comparing across the different HRFs, apparent differences arise in the overall magnitude of effects (far lower betas for the Dale & Buckner HRF), rank ordering of effects (non-repeated body parts vs. repeated faces, for example), and variance explained (Dale & Buckner does far worse). The magnitude of adaptation effects is much greater for some HRFs than others.
Deconvolution explains far more variance than any of the fitted HRFs (35% vs. 19%), and as we can see from the average time course, seems to model a very clean HRF, in which the rise, peak, and undershoot occur at the same time for each trial.
Multi-Voxel Patterns
The multi-voxel patterns validate and extend the GLM results. In the face-selective ROIs, we see more correlated activation between non-repeated and repeated face trials, but no correlations between face patterns and other patterns. Across the other ROIs, we see similar specific adaptation effects for body parts (face-selective ROI #2 and body part-selective ROI) and cars (face-selective ROI #2). We also see some mixed evidence of common patterns of activation across non-repeated categories as shown by the weaker correlations between body parts, cars, and houses.
General Conclusions
In this exploratory study, I examined several different ways to fit predictors to an ROI timecourse and their effects on predictor magnitudes, variance explained, and rank ordering of activations. Moreover, I attempted to understand the data on a finer spatial scale by scrutinizing multi-voxel patterns. While several more analyses would be necessary to test the statistical and pragmatic significance of the adaptation effects found here, this project begins to demonstrate the importance of choices made in the stream of data analysis. fMRI datasets are rich, complex, noisy, multi-componential, and vast. The same dataset can tell very different stories depending on the types of analyses performed and extracting that story is both responsibility of and opportunity for the researcher.
References - Resources and related work
[1] Repetition and the brain: neural models of stimulus specific effects. Kalanit Grill-Spector, Rik Henson and Alex Martin. Trends in cognitive science. 2006 Jan;10(1):14-23. Epub Nov. 28th, 2005.
[2] fMRI-adaptation and category selectivity in human ventral temporal cortex: Regional differences across time scales. Weiner, KS., Sayres, R., Vinberg J., Grill-Spector, K. J Neurophysiol. 2010 Apr 7. [Epub ahead of print] 103(6):3349-3365