Hi, I am pretty new to MR analysis and it is the first time I need to preprocess and analyse multi band sequences. I’ve made basic search through literature, but there wasn’t much on practical approach to the functional MB preprocessing (what, where, how & why). Can you, as least point me to right resources that could guide me though it?
Wishing you all very pleasant weekend!
You can think of typical 2D echo-planar imaging data (fMRI, DWI, etc) as a loaf of sliced bread: a 3D volume is composed of a series of 2D slices. With conventional 2D EPI we acquire one slice at a time, typically in an ascending (1,2,3,…n), descending (n, n-1,…) or interleaved order (1,3,5,… 2,4,6). With multi-band, we acquire multiple slices at the same time. Consider a 6 slice volume with multiband of 2 (e.g. we acquire pairs together). The slice acquisition might be ([1,4],[2,5],[3,6]). The trick is to acquire spatially distant slices simultaneously, and leverage the fact that we are using multiple spatially separate antennas (coils) to acquire the data. Since each antenna has different spatial sensitivity we can reconstruct an image of each slice.
From an analysis perspective you can think of a multiband sequence as a conventional sequence with a very short TR. There are several things to bear in mind:
- When setting up your multiband sequence, follow a few rules:
Make sure that when the number of slices acquired in your volume divided by the multiband factor is an odd number. E.G. For multi-band x2 you should acquire volumes of 34 or 38 slices but should avoid 36 (for details, Figure 3 of Barth et al.).
Limit multi-band factors to minimize aliasing, which is especially problematic for low-dimension coils (e.g. 12 and 20 channel data).
In-plane acceleration (iPAT, SENSE, GRAPPA) also leverage multiple channels. Therefore, as multi-band is increased, one must consider reducing these values (or suffer aliasing artifacts from each method). Therefore, MB sequences typically have low iPAT factors, and consequently have a high amount of EPI artifacts related to long readout times (spatial distortion and signal dropout). This may make MB less suitable if you are interested in regions with high susceptibility (orbital frontal cortex, medial temporal lobes). While the dropout can not be salvaged, the spatial distortion can be mitigated by acquiring undistortion images (either fieldmap for FUGUE or phase-reversed spin echo for TOPUP).
- MB dramatically reduces the repetition time. For MB=3, the TR will be around 1/3 of the time relative to an unaccelerated sequence with x3 the amount of data. Therefore, you have more total signal to noise across the time series and your Nyquist is reduced. However, each image will have reduced signal-to-noise due to the short TR (this reflects T1 effects, with shorter TR there is less time for realignment with the magnetic field).
- While you will acquire dramatically more images, each image will be noisier (due to T1 effects) and the auto-correlation with neighboring scans will be higher. Therefore, while your statistical power will likely increase with MB, the impact will be somewhat mitigated.
- You will typically want to acquire a “single-band reference” image - which is acquired without the aliasing artifacts of MB or the very short TR. In all other aspects, this is the same as the MB image (e.g. same spatial position, slices, iPAT, etc). This image will have much better signal to noise and better tissue contrast. The tissue contrast will make this better for tissue matching (e.g. FSL BBR) than the mean MB series.
- The slice time correction must be either skipped or be aware of the actual slice acquisition order. Slice time correction deals with the fact that we do not see the entire 3D volume at one moment in time, but rather as a series of slices. A major benefit of MB is that the slice time error is much smaller. In many situations, the slice time error could be so small one could simply skip STC for high multi-band factors (error is small relative to HRF).
In general, MB has terrific benefits. The shorter TR dramatically reduces the Nyquist (useful for resting state), one can acquire high-direction, multi-b-value DWI in a reasonable time, one can acquire a lot more fMRI observations in a given time.
This is very extensive and great reply! I do intend to use fmriprep to preprocess the data, but now it makes so much difference with your background. Thank you!
Hi @Chris_Rorden @PuddleJumper, Can you suggest what changes we need to make while running fmriprep command? Do we need to turn on/off any flags?
Hi, The only two flags I used were
--fs-no-reconall --ignore slicetiming
first is for brain surface reconstruction (done by freesurfer, I believe and takes a lot of time) which I didn’t need and second is to ignore slice timing (and that’s a must do for MB sequence) as Chris mentioned.
Just want to mention that using recon-all enables fMRIprep to use bbregister to do surface-based registration instead of volumetric registration, which may perform better. So if you aren’t in a time crunch, you may want to run recon-all.
Two limitations to be aware of with the current fmriprep version (up to v20.2.3):
For the time being, using recon-all creates aberrant TPMs (tissue probabilities map) an issue that is problematic for us. However the great developper team is working on this and a solution will be implemented in the next versions of fmriprep:
More details here:
We noticed an issue when using SBref images: the realignment parameters calculated by mcflirt were somewhat off (many zeros). We currently pass
--ignore sbref flag on our fmriprep executions (option enabled in fmriprep from version 20.2.2) to get more reliable motion parameters with fmriprep.
More details here: