help > Multiband sequence / subsecond TRs
Showing 1-6 of 6 posts
Mar 30, 2020 06:03 PM | Maria Athanassiou - University of Montreal
Multiband sequence / subsecond TRs
Good afternoon everyone,
I'm about to run several analyses of fMRI data that was acquired with multiband sequences. The TR was shorter than 1 second in all cases (0.7 / 0.8 sec). While the advantages of short TRs seem more straightforward for brain activations studies, the advantages do not seem to be so obvious in the case resting-state functional connectivity (rsFC) studies (Chen et al., 2019). I would dearly appreciate to have your valuable input on the potential changes that should be made to the pre- and post-processing of rsFC data of studies using short TRs.
1) Should we still perform slice-time correction ? My understanding is that it would not improve preprocessing.
2) Should we perform particular adjustments to the low-pass filter for rsFC data acquired using subsecond TRs ?
3) For studies on task-based activations, it is recommended to pay careful attention on the choice of the method used to correct for temporal correlations for studies using multiband sequences (e.g. AR2 and FAST rather than AR1, for instance). Is there a similar issue in the case of resting-state functional connectivity data ? How is CONN handling this issue ?
4) Do you have any other recommendations ?
Many thanks for your attention,
Maria
Reference: Chen J, Polimeni JR, Bollmann S, Glover GH. On the analysis of rapidly sampled fMRI. Neuroimage 2019; 188: 807-820.
I'm about to run several analyses of fMRI data that was acquired with multiband sequences. The TR was shorter than 1 second in all cases (0.7 / 0.8 sec). While the advantages of short TRs seem more straightforward for brain activations studies, the advantages do not seem to be so obvious in the case resting-state functional connectivity (rsFC) studies (Chen et al., 2019). I would dearly appreciate to have your valuable input on the potential changes that should be made to the pre- and post-processing of rsFC data of studies using short TRs.
1) Should we still perform slice-time correction ? My understanding is that it would not improve preprocessing.
2) Should we perform particular adjustments to the low-pass filter for rsFC data acquired using subsecond TRs ?
3) For studies on task-based activations, it is recommended to pay careful attention on the choice of the method used to correct for temporal correlations for studies using multiband sequences (e.g. AR2 and FAST rather than AR1, for instance). Is there a similar issue in the case of resting-state functional connectivity data ? How is CONN handling this issue ?
4) Do you have any other recommendations ?
Many thanks for your attention,
Maria
Reference: Chen J, Polimeni JR, Bollmann S, Glover GH. On the analysis of rapidly sampled fMRI. Neuroimage 2019; 188: 807-820.
Apr 1, 2020 09:04 PM | Alfonso Nieto-Castanon - Boston University
RE: Multiband sequence / subsecond TRs
Hi Maria,
Regarding (1), for TR below 1s and particularly if you are mainly interested in low-frequency fluctuations (below 0.1Hz) I agree that skipping slice-timing correction is perfectly reasonable (but this would not be the case if you are interested in high-frequency fluctuations and/or if you have a fast event-related design, where small time differences between slices may still play a significant role)
Regarding (2), fast acquisitions allow you to evaluate BOLD signal fluctuations at higher frequencies (e.g. and study possible frequency dependency of effects of interest). In those scenario you may want to increase the low-pass filter threshold -e.g. from 0.1Hz to 0.2Hz- in order to keep some of those higher frequencies into your analyses. That said, keep in mind that respiratory and aliased cardiac effects may also be now more visible/focused at around 0.3Hz, so you will often want to filter those higher frequencies out of the BOLD signal (as fast acquisitions increase the effectiveness of frequency-filtering as a denoising strategy).
Regarding (3), as far as I am aware those issues are only relevant for the accurate estimation of single-subject statistics, as they affect the strength and form of the temporal autocorrelation in the BOLD signal, but they do not affect single-subject effect-size estimation or group-level analyses, where measures such as functional connectivity are generally invariant to the structure of the BOLD signal temporal autocorrelation and are never "corrected" by the estimated effective degrees of freedom of the BOLD timeseries (see for example this post for a relevant reference https://www.nitrc.org/forum/message.php?...).
And regarding (4), I would simply recommend as always to spend some time evaluating and trying to improve the properties of the BOLD signal and the voxel-to-voxel correlation distributions after denoising using the corresponding Quality Assurance metrics/plots (in CONN these are all of the QA plots with labels "QA denoising: ..."), since "optimal" preprocessing and denoising procedures may well vary across different studies and samples, so all of the above general recommendations should be taken only as a reasonable starting point for your analyses.
Hope this helps
Alfonso
Originally posted by Maria Athanassiou:
Regarding (1), for TR below 1s and particularly if you are mainly interested in low-frequency fluctuations (below 0.1Hz) I agree that skipping slice-timing correction is perfectly reasonable (but this would not be the case if you are interested in high-frequency fluctuations and/or if you have a fast event-related design, where small time differences between slices may still play a significant role)
Regarding (2), fast acquisitions allow you to evaluate BOLD signal fluctuations at higher frequencies (e.g. and study possible frequency dependency of effects of interest). In those scenario you may want to increase the low-pass filter threshold -e.g. from 0.1Hz to 0.2Hz- in order to keep some of those higher frequencies into your analyses. That said, keep in mind that respiratory and aliased cardiac effects may also be now more visible/focused at around 0.3Hz, so you will often want to filter those higher frequencies out of the BOLD signal (as fast acquisitions increase the effectiveness of frequency-filtering as a denoising strategy).
Regarding (3), as far as I am aware those issues are only relevant for the accurate estimation of single-subject statistics, as they affect the strength and form of the temporal autocorrelation in the BOLD signal, but they do not affect single-subject effect-size estimation or group-level analyses, where measures such as functional connectivity are generally invariant to the structure of the BOLD signal temporal autocorrelation and are never "corrected" by the estimated effective degrees of freedom of the BOLD timeseries (see for example this post for a relevant reference https://www.nitrc.org/forum/message.php?...).
And regarding (4), I would simply recommend as always to spend some time evaluating and trying to improve the properties of the BOLD signal and the voxel-to-voxel correlation distributions after denoising using the corresponding Quality Assurance metrics/plots (in CONN these are all of the QA plots with labels "QA denoising: ..."), since "optimal" preprocessing and denoising procedures may well vary across different studies and samples, so all of the above general recommendations should be taken only as a reasonable starting point for your analyses.
Hope this helps
Alfonso
Originally posted by Maria Athanassiou:
Good afternoon everyone,
I'm about to run several analyses of fMRI data that was acquired with multiband sequences. The TR was shorter than 1 second in all cases (0.7 / 0.8 sec). While the advantages of short TRs seem more straightforward for brain activations studies, the advantages do not seem to be so obvious in the case resting-state functional connectivity (rsFC) studies (Chen et al., 2019). I would dearly appreciate to have your valuable input on the potential changes that should be made to the pre- and post-processing of rsFC data of studies using short TRs.
1) Should we still perform slice-time correction ? My understanding is that it would not improve preprocessing.
2) Should we perform particular adjustments to the low-pass filter for rsFC data acquired using subsecond TRs ?
3) For studies on task-based activations, it is recommended to pay careful attention on the choice of the method used to correct for temporal correlations for studies using multiband sequences (e.g. AR2 and FAST rather than AR1, for instance). Is there a similar issue in the case of resting-state functional connectivity data ? How is CONN handling this issue ?
4) Do you have any other recommendations ?
Many thanks for your attention,
Maria
Reference: Chen J, Polimeni JR, Bollmann S, Glover GH. On the analysis of rapidly sampled fMRI. Neuroimage 2019; 188: 807-820.
I'm about to run several analyses of fMRI data that was acquired with multiband sequences. The TR was shorter than 1 second in all cases (0.7 / 0.8 sec). While the advantages of short TRs seem more straightforward for brain activations studies, the advantages do not seem to be so obvious in the case resting-state functional connectivity (rsFC) studies (Chen et al., 2019). I would dearly appreciate to have your valuable input on the potential changes that should be made to the pre- and post-processing of rsFC data of studies using short TRs.
1) Should we still perform slice-time correction ? My understanding is that it would not improve preprocessing.
2) Should we perform particular adjustments to the low-pass filter for rsFC data acquired using subsecond TRs ?
3) For studies on task-based activations, it is recommended to pay careful attention on the choice of the method used to correct for temporal correlations for studies using multiband sequences (e.g. AR2 and FAST rather than AR1, for instance). Is there a similar issue in the case of resting-state functional connectivity data ? How is CONN handling this issue ?
4) Do you have any other recommendations ?
Many thanks for your attention,
Maria
Reference: Chen J, Polimeni JR, Bollmann S, Glover GH. On the analysis of rapidly sampled fMRI. Neuroimage 2019; 188: 807-820.
Apr 9, 2020 05:04 PM | Maria Athanassiou - University of Montreal
RE: Multiband sequence / subsecond TRs
Dear Alfonso,
Many thanks for your helfpul answer!
We have one last question, regarding the use of a multiband sequence for an fMRI study using a rapid event-related paradigm.
We are trying to use the confound-corrected time series from CONN in SPM12 for the first-level analysis. 42 slices ares acquired, the TR=0.785, and the multi-band sequence (factor = 3) is the following:
547.5 0 382.5 55 437.5 110 492.5 217.5 600 272.5 655 327.5 710 165 547.5 0 382.5 55 437.5 110 492.5 217.5 600 272.5 655 327.5 710 165 547.5 0 382.5 55 437.5 110 492.5 217.5 600 272.5 655 327.5 710 165
In SPM-12, for the first-level analysis, we have to specify the microtime resolution and microtime onset. Do we still use 42 slices as the microtime resolution and 21 slices as the microtime onset ? Or should we make adjustments ?
Thank your again for your help,
Maria
Many thanks for your helfpul answer!
We have one last question, regarding the use of a multiband sequence for an fMRI study using a rapid event-related paradigm.
We are trying to use the confound-corrected time series from CONN in SPM12 for the first-level analysis. 42 slices ares acquired, the TR=0.785, and the multi-band sequence (factor = 3) is the following:
547.5 0 382.5 55 437.5 110 492.5 217.5 600 272.5 655 327.5 710 165 547.5 0 382.5 55 437.5 110 492.5 217.5 600 272.5 655 327.5 710 165 547.5 0 382.5 55 437.5 110 492.5 217.5 600 272.5 655 327.5 710 165
In SPM-12, for the first-level analysis, we have to specify the microtime resolution and microtime onset. Do we still use 42 slices as the microtime resolution and 21 slices as the microtime onset ? Or should we make adjustments ?
Thank your again for your help,
Maria
Apr 10, 2020 07:04 PM | Alfonso Nieto-Castanon - Boston University
RE: Multiband sequence / subsecond TRs
Dear Maria,
Yes, that's exactly correct, CONN&SPM slice-timing correction uses as reference the mid-TR slice, so in your SPM first-level analyses you may simply use the default settings (microtime resolution 16, microtime onset 8) or you may change that to 42 and 21, both would be perfectly fine (just for reference, this only applies to SPM12; if I recall correctly in SPM8 and below there was an inconsistency in the microtime onset default value -1st bin, unlike SPM12- and the slice-timing correction reference slice -mid slice just as SPM12-)
Best
Alfonso
Originally posted by Maria Athanassiou:
Yes, that's exactly correct, CONN&SPM slice-timing correction uses as reference the mid-TR slice, so in your SPM first-level analyses you may simply use the default settings (microtime resolution 16, microtime onset 8) or you may change that to 42 and 21, both would be perfectly fine (just for reference, this only applies to SPM12; if I recall correctly in SPM8 and below there was an inconsistency in the microtime onset default value -1st bin, unlike SPM12- and the slice-timing correction reference slice -mid slice just as SPM12-)
Best
Alfonso
Originally posted by Maria Athanassiou:
Dear Alfonso,
Many thanks for your helfpul answer!
We have one last question, regarding the use of a multiband sequence for an fMRI study using a rapid event-related paradigm.
We are trying to use the confound-corrected time series from CONN in SPM12 for the first-level analysis. 42 slices ares acquired, the TR=0.785, and the multi-band sequence (factor = 3) is the following:
547.5 0 382.5 55 437.5 110 492.5 217.5 600 272.5 655 327.5 710 165 547.5 0 382.5 55 437.5 110 492.5 217.5 600 272.5 655 327.5 710 165 547.5 0 382.5 55 437.5 110 492.5 217.5 600 272.5 655 327.5 710 165
In SPM-12, for the first-level analysis, we have to specify the microtime resolution and microtime onset. Do we still use 42 slices as the microtime resolution and 21 slices as the microtime onset ? Or should we make adjustments ?
Thank your again for your help,
Maria
Many thanks for your helfpul answer!
We have one last question, regarding the use of a multiband sequence for an fMRI study using a rapid event-related paradigm.
We are trying to use the confound-corrected time series from CONN in SPM12 for the first-level analysis. 42 slices ares acquired, the TR=0.785, and the multi-band sequence (factor = 3) is the following:
547.5 0 382.5 55 437.5 110 492.5 217.5 600 272.5 655 327.5 710 165 547.5 0 382.5 55 437.5 110 492.5 217.5 600 272.5 655 327.5 710 165 547.5 0 382.5 55 437.5 110 492.5 217.5 600 272.5 655 327.5 710 165
In SPM-12, for the first-level analysis, we have to specify the microtime resolution and microtime onset. Do we still use 42 slices as the microtime resolution and 21 slices as the microtime onset ? Or should we make adjustments ?
Thank your again for your help,
Maria
Apr 16, 2020 09:04 PM | Maria Athanassiou - University of Montreal
RE: Multiband sequence / subsecond TRs
Hi Alfonso,
Thank you very much for the clarifications, it is greatly appreciated!
I would have a small side question. I'm coming across an additional issue that I would like your input on.
I'm using my niftiDATA files (output confound corrected time series of CONN) for my first level analyses, specifying my conditions, and conducting a model estimation for each subject. Yet, when I view the results, I'm getting values above 100, when I should be getting standardized values between -1 and 1. Would you happen to have any insight as to why this may be occurring?
Thank you very much in advance, and have a lovely evening,
Maria
Thank you very much for the clarifications, it is greatly appreciated!
I would have a small side question. I'm coming across an additional issue that I would like your input on.
I'm using my niftiDATA files (output confound corrected time series of CONN) for my first level analyses, specifying my conditions, and conducting a model estimation for each subject. Yet, when I view the results, I'm getting values above 100, when I should be getting standardized values between -1 and 1. Would you happen to have any insight as to why this may be occurring?
Thank you very much in advance, and have a lovely evening,
Maria
Apr 18, 2020 04:04 PM | Alfonso Nieto-Castanon - Boston University
RE: Multiband sequence / subsecond TRs
Hi Maria,
It depends on software details but, for example, if you are running first-level analyses using SPM, this could arise from SPM default "grand mean scaling" behavior, which scales all effect-sizes to percent signal change values (estimated with respect to the average BOLD signal within a brainmask). The problem with this scaling is that nifti* files will typically have the average BOLD signal removed from the timeseries as part of the default denoising procedure in CONN, so the reference average BOLD signal becomes meaningless and the effect-sizes very-strangely scaled.
If that is the case here, to avoid that I would recommend simply to use CONN's modular commands for denoising as described in https://web.conn-toolbox.org/fmri-method... which will create a different series of denoised d*.nii files (instead of your current nifti*.nii files). This procedure is the one I recommended for denoising your functional data for non-connectivity analyses, and it has a few differences with the standard CONN denoising pipeline because of that (one of the differences is that it will automatically re-enter the BOLD average into your denoised dataset precisely to avoid this sort of scaling issues with SPM; another difference is that it will not use automatically the "effect of task" regressors as part of denoising since those are often part of your effects of interest for functional activation analyses)
Hope this helps
Alfonso
Originally posted by Maria Athanassiou:
It depends on software details but, for example, if you are running first-level analyses using SPM, this could arise from SPM default "grand mean scaling" behavior, which scales all effect-sizes to percent signal change values (estimated with respect to the average BOLD signal within a brainmask). The problem with this scaling is that nifti* files will typically have the average BOLD signal removed from the timeseries as part of the default denoising procedure in CONN, so the reference average BOLD signal becomes meaningless and the effect-sizes very-strangely scaled.
If that is the case here, to avoid that I would recommend simply to use CONN's modular commands for denoising as described in https://web.conn-toolbox.org/fmri-method... which will create a different series of denoised d*.nii files (instead of your current nifti*.nii files). This procedure is the one I recommended for denoising your functional data for non-connectivity analyses, and it has a few differences with the standard CONN denoising pipeline because of that (one of the differences is that it will automatically re-enter the BOLD average into your denoised dataset precisely to avoid this sort of scaling issues with SPM; another difference is that it will not use automatically the "effect of task" regressors as part of denoising since those are often part of your effects of interest for functional activation analyses)
Hope this helps
Alfonso
Originally posted by Maria Athanassiou:
Hi Alfonso,
Thank you very much for the clarifications, it is greatly appreciated!
I would have a small side question. I'm coming across an additional issue that I would like your input on.
I'm using my niftiDATA files (output confound corrected time series of CONN) for my first level analyses, specifying my conditions, and conducting a model estimation for each subject. Yet, when I view the results, I'm getting values above 100, when I should be getting standardized values between -1 and 1. Would you happen to have any insight as to why this may be occurring?
Thank you very much in advance, and have a lovely evening,
Maria
Thank you very much for the clarifications, it is greatly appreciated!
I would have a small side question. I'm coming across an additional issue that I would like your input on.
I'm using my niftiDATA files (output confound corrected time series of CONN) for my first level analyses, specifying my conditions, and conducting a model estimation for each subject. Yet, when I view the results, I'm getting values above 100, when I should be getting standardized values between -1 and 1. Would you happen to have any insight as to why this may be occurring?
Thank you very much in advance, and have a lovely evening,
Maria