Scanning FAQ

Frequently Asked Questions - fMRI Scanning

This section is intended to address design-related questions that focus primarily on technical aspects about the scanner - things like TR, pulse sequence, slice thickness, etc. Obviously, setting your scanner parameters is mixed in heavily with your experimental design, so be sure to check out some other design-related pages:

ScanningPapers has some nice handouts from Gary Glover and Philippe, summarizing some of these articles and (very informally and clearly) addressing questions about TR length, signal-to-noise tradeoffs, what k-space is exactly, etc. Definitely check those out.

1. What pulse sequence shoudl I use (EPI or spiral)?: What are pros and cons of each? What do each of them get you?

  • EPI: More widely used, and hence supported by all fMRI analysis programs. Some programs (FSL, or SPM's unwarping module, for example) do not support spiral data. Can be subject to less drop-out in some regions than spiral-in or spiral-out data alone. Can be easier to figure out what the slice ordering is.
  • Spiral: Properly weighted and combined, spiral in-out shows significantly less signal drop-out and shows significantly greater activations in many areas of the brain, including ventromedial PFC, medial temporal lobe, etc. Effect is even more pronounced at higher field strengths (see Preston et. al on ScanningPapers). However, it is less widely supported, and Gary's trademark spiral i-o sequence may not even be physically possible on some other institutions' equipment.

2. What should your TR be? What are the tradeoffs, and what's the best tradeoff of coverage vs. speed for different types of analysis?

Bottom line: TR should be as short as possible, given how many slices you want to cover and the limits of your task. Gary's handout and monograph on ScanningPapers speak best to this issue, and are good quick reads. Decreasing your TR decreases your signal-to-noise ratio (SNR) in any one functional image, but because you have more images to work with, your overall SNR increases with decreasing SNR. Your TR, however, is limited by how many slices you want to take. On the 3T scanner here at Stanford, using spiral in-out, each slice takes approximately 65 msec/slice (TE = 30 msec), so you can get 15 full slices in 1 second (on the 1.5T, slices take about 75 msec, so you can get 13 full slices in 1 sec.). Your tradeoff is that with fewer slices, you have to either accept less coverage of the brain, or thicker slices, which will have poorer resolution in the z direction.

For certain experiments, then - ones focused on primary sensory on motor cortices, when you don't care about the rest of the brain - you can buy yourself shorter TRs by decreasing your number of slices, or you can increase your number of slices and make them smaller, while keeping your TR constant. Assuming you need full coverage of the brain, you can only decrease your TR by making your slices thicker, which you should do as much as possible within the constraints of your desired resolution.

Alternatively, in experimental designs where you're not particularly focused on timecourse information and where you already have good statistical power - namely, block-design experiments - you may want to get better resolution by increasing your number of slices and hence your TR. In event-related experiments, having a short TR becomes even more imporant, due to the relative lack of experimental power in such designs and relative importance of timecourse information.

3. What should your slice thickness be?

As thick as you think you can get away with. Increasing your slice thickness allows you to decrease your TR and maintain the same coverage, which is desirable as you get better SNR with decreasing TR. Alternatively, if you need good resolution in all dimensions, you can shrink your slice thickness at the expense of either brain coverage or having a longer TR.

4. What should your slice resolution / voxel size be?

64 x 64 is standard around here for full-brain coverage. With experiments focusing on smaller areas - primary motor and/or sensory cortices - something else (like 128 x 128) may be useful to get better resolution in a smaller area.

This differs from what size you interpolate your voxels to in normalization, which is covered in NormalizationFaq...

5. Should you acquire axially/coronally/something else? How come?

Big issue here, as I understand it, is that your slices are often thicker than your in-slice voxels, and hence your resolution is often poorest in the direction perpendicular to your slices. (Hence, if you acquire axially, your inferior-superior or z-direction resolution may not be great.) If you have a particular structure of interest, depending on its orientation, you may want to arrange your slices so as to get good resolution in the direction necessary to nail down that structure. Anyone else have any comments on this one?

6. BOLD vs. perfusion: what are pros and cons of each? What sorts of experiments would you use perfusion for?

Perfusion imaging - in which arterial blood is magnetically 'labeled' with an RF pulse, and then tracked as it moves through the brain - has two main advantages we discussed, one of which is thoroughly discussed in the Aguirre article in ScanningPapers. That advantage is the relatively different noise profile present within the perfusion signal. Unlike BOLD, perfusion noise doesn't have very much autocorrelation, which isn't by itself anything special, but means that perfusion contains much, much less noise relative to BOLD at very low experimental frequencies. There is more noise in general, though, in perfusion imaging, so in general SNR ratios are better for BOLD. But for experiments with very low task-switching frequency - say, blocks of 60s or more, even up to many minutes or hours - BOLD is almost useless, due to the preponderance of low-frequency noise, whereas the perfusion signal is unchanged. This means that with block lengths of longer than a minute, perfusion imaging is probably a better way to go, and experiment which previously weren't possible - block lengths of several minutes, or task switching taking place over several days - might be designed with perfusion.

Another feature of the noise in perfusion imaging is that it appears to be more reliable across subjects. While SNR within a given subject is higher for BOLD, group SNR appears (with limited data in Aguirre on ScanningPapers) to be higher in perfusion imaging. More research is needed on this subject, but this relative SNR advantage may be useful for experiments with small numbers of subjects, as across-subject variability is always the largest noise source for BOLD experiments, often by a huge factor.

The other primary advantage of perfusion relative to BOLD imaging is that the perfusion signal is an absolute number, rather than a contrast. Each voxel is given a physiologically intelligible value - amount of cerebral blood flow - which means that it can be especially useful for comparing groups of populations. Comparing the results of a particular contrast in depressed vs. normal subjects might not yield any results, for example, but overall blood flow might just be lower in the absolute in a particular regions for depressed subjects relative to normal subjects - which would be very interesting. The ability to compare baselines in perfusion is a strong case for using it in particular types of experiments where baseline information may be interesting.