Education, tips and tricks to help you conduct better fMRI experiments.
Sure, you can try to fix it during data processing, but you're usually better off fixing the acquisition!

Wednesday, December 22, 2010

Resting state fMRI - part III

The story so far...

Finally, here is the third part of a three-part series of posts that have sought to determine a general protocol for resting state fMRI (rs-fMRI). In the first post I reviewed a paper by van Dijk et al. that showed that spatial and temporal resolution didn't make a huge difference to the way resting state networks could be detected using current methods (i.e. seed cross correlation or ICA).

In the second post I presented the results of some simple tests that aimed to determine what sort of spatial coverage could be attained with parameters in accordance with the conclusions of the van Dijk paper. Temporal SNR (TSNR) was used as a simple proxy for data quality. It was found that TSNR for 3.5 mm in-plane resolution was fairly consistent across a range of axial and axial-oblique slice orientations, as well as for sagittal slices.

One question remained, however: given the tolerance to a longish TR (compared to event-related fMRI) for detecting resting networks, would it be beneficial to acquire many thinner slices in a longer TR, or fewer thicker slices in a shorter TR? Following van Dijk et al. we wouldn't expect any huge penalty from extending the TR a bit, but there might be a gain of signal in regions suffering extensive dropout which would suggest that thinner slices might be useful.

The final experiment

Which brings us to today's post. The final comparison is between as many slices as we can get in a TR of 2500 ms, versus as many thinner slices as we can get in a TR of 3000 ms. Obviously, if one is going to increase the TR by a factor of 1.2 it follows that the slice thickness can be reduced by a similar factor to produce the same 3D volume coverage, ignoring a small discrepancy in inter-slice gap. Thus, the specific comparison was between:

(a) TR=2500 ms, slice thickness = 3 mm, gap = 0.3 mm, 43 interleaved slices, number of volumes = 144.

(b) TR=3000 ms, slice thickness = 2.5 mm, gap = 0.25 mm, 52 interleaved slices, number of volumes = 120.

The overall acquisition duration was 6 minutes for each test. Other acquisition parameters were:

Siemens 3 T Trio/TIM running VB15, 12-channel HEAD MATRIX coil, ep2d_bold pulse sequence, matrix=64x64, FOV=224x224 mm, TE=25 ms, bandwidth=2056 Hz/pixel, echo spacing=0.55 ms, fatsat=ON, MoCo=ON, no spatial filters.

A single slice orientation was tested: an axial-oblique prescription tilted approximately 15 degrees towards coronal, for efficient brain coverage. The flip angle was left at 78 degrees for the two experiments even though the TR was altered. (Indeed, the chances are a 90 degree flip angle could have been used for both TRs, so the 78 degree flip was considered to be valid for a direct comparison.) The subject - me again - lay in the magnet staring at the blue stripe on the magnet bore and tried (valiantly) not to fall asleep. (Is it too late to patent MRI scanners as a cure for insomnia...?)


TSNR images were produced for each image series (raw and rigid body motion-corrected), and regions of interest were compared. Interestingly, the 3 mm slices produced TSNR only 6% higher than the 2.5 mm slices. Neglecting physiologic noise and all that good stuff for a moment, based on voxel volume alone one would expect 20% higher (static) image SNR. And with an additional 24 frames (144 vs 120) a further 9.5% benefit to (static) image SNR should arise. The near 30% gain in (static) SNR didn't translate into such large gains in TSNR.

Though no attempt was made to quantify the reduced dropout, there didn't appear to be any significant effect of the thinner slices in the temporal and frontal lobes. Importantly there were no obvious artifacts in either series; the thinner slices didn't introduce any negative consequences visible on the TSNR images. Indeed, the general features of the 2.5 and 3 mm slice TSNR were remarkably consistent, although the edges of the brain in the thinner slices were marginally 'sharper' than for the 3 mm slices.

Conclusions and recommendations

Using 20% thinner slices, a proportionally longer TR and a fixed scan duration of 6 minutes produced only a 6% hit to TSNR, yet with a small (unquantified) improvement in the signal at the edges of the brain and possibly a minor benefit to high susceptibility regions. 

So what does all this mean to you? In a nutshell, most of these parameters are a wash. Using a TR between 2500 and 3000 ms to cover the entire brain as efficiently as possible (i.e. thinnest slices, largest matrix in the time allotted) seems to be quite robust to small changes in any spatial parameter as measured by TSNR.

What, then, should we consider to be "standard," if anything? Assuming processing methods such as seeded cross correlations or ICA, and the aim to characterize what are presently considered to be the "typical" resting networks (DMN, etc.), then the following rough parameters appear to be suitable:

- TR in the range 2500-3000 ms.

- FOV in the range 192-224 mm.

- TE in the range 23-30 ms.

- Axial oblique slices in the range 10-30 degrees towards the coronal plane (though sagittal could be a viable option if brain stem is of interest).

It all looks quite generic, doesn't it? And I suppose that's the point. If there is no major advantage from a TR of 2500 ms versus 3000 ms, or an in-plane resolution of 3 mm versus 2.5 mm, then we might as well pick a set of numbers and stick with them. That way, if lots of people are using the same set of numbers it might make for easier comparisons between studies as well as opportunities for meta-analysis of pooled data. (Hey, we can all dream!)

Note, of course, that the tests presented in this series of posts are disposable: I was simply looking for show-stopping parameters, not an exhaustive hunt for the perfect protocol. But it does seem that there is robustness to the actual parameter set. Thus, making small changes to the above parameters is not expected to have large consequences (positive or negative).

Going out on a limb

Here goes nothing. I'm going to pick a set of numbers and await your opinions. I'm going with the small benefit to TSNR of the 3 mm slices, and the possibility that respiration effects might just be properly sampled (Nyquist criterion) with a TR of 2500 ms. And although I'm going to recommend an axial-oblique slice prescription for easy coverage of the entire cortex and cerebellum, I'd be interested to see what people get with sagittal slices. Thus, the final list of parameters:

TR=2500 ms, TE=25 ms, FOV=224x224 mm, matrix=64x64, forty-three 3 mm axial-oblique slices (10% gap) aligned approx 15 degrees towards coronal, flip angle=90 deg, 144 volumes (for a 6 minute run).

Physiological monitoring - heart rate and respiration - are also strongly advised, athough the benefit of these measures is still under debate. The way I see it, if you acquire the physio data you will have reserved the option to use them should a subsequent study find significant benefits. If you don't, well, you can figure out what you can't do!



Want the data from this post?

You can download a zip file containing all the raw DICOM images as well as DICOM versions of the mean, stdev and TSNR images here:

If you don’t already have a DICOM viewer, check out Osirix for Mac OSX (available via a link in the side bar). ImageJ from NIH also has some nice features for ROI analysis. I’ll post introductions to using these two programs in future posts.


  1. Well, one question is the gap. Why use it and not do a descending interleaved with no gap?
    Thank you for the post, very interesting!



  2. @Anonymous:

    Yep, descending or ascending would work (and is probably better in the presence of significant short-lived motion). But then you'd need between 5 and 20% gap (percent of slice thickness) to avoid cross-talk between adjacent slices. (In some tests I did ages ago I found that only 0% gap actually caused significant cross-talk.) Why the range of gap? Because of the side lobes and general trapezoidal nature of the slice profiles. Turns out it doesn't make that much difference for 2-4 mm slices.

    In sum, then, if you wanted the very best possible setting I'd suggest a 10% gap with descending or ascending slices. I used interleaved in these tests purely because I didn't remember to change it! (And in the absence of acute motion, which can cause stripes in the slice dimension with interleaved slices due to perturbation of the T1 steady state, it doesn't matter that much. I was the subject and I was careful not to move!)

  3. Ben, thank you for the response. Yes true, for motion purposes descending is better, but because I am trying to go with no gap (I'm looking at thalamus, so resolution is precious..) interleaved seems the only way to go to avoid significant cross-talk (though, yes at the risk of big stripes in case of motion during a TR -- why do people not move only in the few msecs in-between TRs??).

    So, for comparison purposes, here is what I am trying to go with (though suggestions & criticisms are welcome):

    TR 3,000 ms (47 slices, 0 gap, interleaved)
    TE 30 ms
    FA 90
    FOV read 192 mm (100% phase; 13% phase oversampling)
    Vox size 3x3x3
    (rel SNR = 1.00)
    Repetitions 200
    Base resolution 64
    Bandwidth 2604 Hz/Px

    Sounds reasonable?



  4. @MCube,

    I'm not at the scanner and I'm away for the next week (ISMRM) so the first question is whether the 2604 Hz/Px bandwidth sits in a zone of mechanical resonance on your Trio. On mine (and others that I am familiar with), for axial or axial-obl slices (where X is read grad) we avoid 0.6-0.8 ms echo spacing to avoid mech resonances and above minimum ghosting.

    My next question: why the oversampling? This is equivalent to using a larger FOV and simply restricting the part of the image you want to view. As a matter of course I would always only acquire data points I want, so to maintain voxel size I would set no oversampling and reduce the phase encode matrix size proportionally. It all comes out in the wash, but as a general rule I don't know of a reason to oversample. It's equivalent (in terms of hardware performance needs) to using a larger FOV and matrix, and cropping the resultant image after FFT.

    Those two points are relatively benign. I think the rest of your params look perfectly reasonable too. If you have any residual concern about the interleaved no gap vs descending 5% gap (say), then I would suggest you acquire a couple of test data sets with the same 200 reps, back to back on a representative subject. Run the two data sets through your offline processing pipeline (especially realignment) as you'll use for your experiment, then generate TSNR maps. Locate the thalamus on either a raw EPI or on an MP-RAGE from the same session, and see which of the two slicing options is better, if at all.

    Re. motion generally, as I just informed another reader offline, tell the subject not to move if he hears the scanner going ping! Scratching nose, wiggling feet, stretching back etc can be done reasonably safely when the scanner is silent. (Re-shim if you want between EPI blocks - see last week's post.) In my experience ANY sort of body motion causes head motion, not just directly moving the head! Our necks will pitch with upper arm movement, leg movement, bum movement...

  5. Hi Ben, I hope Montreal was fun!

    So, thank you for your suggestions. I think the oversampling ended up being the way to get the TR within 3s and the 3mm^3 resolution (and the matrix res above 64). I will go back to the hospital tomorrow so I should be able to play with the parameters a little more. If I get rid of the oversampling, won't I have to trade that off for a longer TR or larger voxels?

    Re the bandwidth, I was wondering about it. I didn't set it -- does the machine (it's an Allegra, not a TTrio) calculate that? If not, then I just inherited it from a previous sequence. I must confess though, I know what the bandwidth is, but I don't think I understand it's "consequences" well enough to find a 'better' value, so I'd like to know if you think there is a better one (or how can I go by finding it out..)

    Let me know what your thought are -- and thank you, I wish there were many more physicists with whom I could talk (and actually understand what they are saying..!!)


  6. @MCube: Ah, Montreal... I'm never drinking again :-|

    Echo Spacing/Bandwidth:

    On the Sequence card (Part 1) the Bandwidth (in Hz/Px) can be set in concert with the Echo spacing parameter. On a Trio you want to avoid 0.6-0.8 ms echo spacing for axial or axial oblique slices because of mechanical resonances in the X read gradient (which is L-R for the magnet). You could use an echo spacing in that range if there was no alternative, but the ghosting will be a bit higher. (I'll mention this factor again when I finally get to ghosting in my physics background series of posts.)

    In general (Trio, Allegra, whatever), use the lowest Bandwidth compatible with the echo spacing you select, because this minimizes the amount of ramp sampling versus sampling data on the flat portions of the read gradient episodes. (This also tends to reduce ghosts.) But, don't set bandwidth too low and make the echo spacing too long or distortion will get worse! For an Allegra I really couldn't say what's ideal, perhaps someone else reading the blog can comment? Failing that, you could run some tests on a phantom and evaluate the ghosts for various combinations of echo spacing and bandwidth, keeping your image matrix constant. If possible I would aim for an echo spacing down towards 0.5-0.56 ms for a 70x70 matrix and 210 mm FOV (for 3x3 mm pixels).


    This is one of those things that comes out in the wash. You're using a small nominal FOV of 192 mm with oversampling. Alternatively, you could set up a larger FOV with proportionately larger matrix size and no oversampling, thereby maintaining the pixel size, and the pulse sequence timing parameters would be similar. You're not really gaining much in terms of performance, and with 3 mm voxels you're nowhere near to the physical limits of the system. In using a smaller FOV and oversampling you save a tiny bit of hard disk space (smaller images), and perhaps make your echo spacing a very small amount shorter, but at the expense of not having a true "noise" region around the head. (My preference is to have a small noise region in which to evaluate ghosts and other artifacts, which can be really hard to evaluate once almost the entire image is full of brain.) If you change the echo spacing and bandwidth appropriately, along with the FOV (larger) and matrix (larger) then you should see no effect on min. TE or on TR.

    Another concern with 192 mm FOV is that large heads and/or slight misplacement of slices could easily lead to you missing a portion of the brain completely! Much less of a risk with a 210 mm FOV. I'd be interested to know what the echo spacing and bandwidth can be set to with 3 mm voxels, FOV 210 mm, 70x70 matrix, no oversampling. I'll see if I can make a quick comparison on my Trio later today, I'm curious to see if there's a significant difference on my scanner now.