High frequency noise related to MB and temporal ICA

Michael Stevens

unread,

Mar 10, 2026, 12:08:21 PMMar 10

to HCP-Users

Our group has been getting some push-back about using higher MB factors with HCP pulse sequences lately... at several levels -- our collaborators raising concerns, at peer review of manuscripts, and even new grant application reviews. There seems to be an increasing concern about the presence of uncorrected high frequency noise that specifically reduces signal detectability in subcortical "center of brain" regions. These concerns have been expressed frequently enough to our group over the past 6 months that we realize it's now "out there" as a real thing we have to carefully address. Now that we're looking closely, we do see some worrisome differences in our newer studies particularly in basic caudate/VS signal detection using identical fMRI paradigms to our older work.

Some colleagues we've spoken with simply advocate dropping the use of MB altogether when the project focus is on subcortical regions. But that essentially makes it difficult or impossible to achieve the spatial resolution needed to optimize HCP processing approaches on our own data. I began to wonder if the temporal ICA denoising approach that the HCP group has been working on for the past few years addresses this issue at all? Or if other, additional denoising steps that are compatible with HCP pipeline-driving processing might be introduced? If tICA is intended to remove or even just mitigate this specific problem, is the automatic application of tICA currently implemented in the most recent HCP pipeline release? And if the answer to that is yes, can I be pointed to an information resource that details implementation... or even a knowledgeable contact who'd be willing to advise? Switching will be non-trivial for us given lots of planned changes in our local cluster configuration... our institution is basically doing a $10M upgrade to a new set of VM- and cloud-based resources... So we have to re-map everything. But the collective worries about this problem has become enough that I'm willing to put the effort into figuring out a whole new approach to data prep at our institution if necessary.

Thanks in advance for any insights or direction folks can offer.

Mike

Glasser, Matthew

unread,

Mar 10, 2026, 2:13:53 PMMar 10

to hcp-...@humanconnectome.org

Hi Mike,

What I currently recommend for fMRI at 3T for new projects (and in progress clinically) is a little different from what the HCP acquired in the past:

*2.4mm isotropic

*TR=1s

*60 angled slices

A Prisma can achieve the above with MB=4. A high-end XT Vida can achieve it with MB=5. A low-end XQ Vida/Skyra requires MB=6. All of these are producing very nice images (better than 3T HCP).

For 7T fMRI, I would stick with what we did with HCP-YA (1.6mm isotropic, TR=1s, 85 angled slices, MB=5, IPAT=2).

The gains of going substantially faster than 1s are not that high in simulations and so in so far as simulations don’t adequately account for all the losses, one might not be gaining any more. Up to 1s is a lot steeper gain. The HCP-Style approach recommends voxel size less than mean cortical thickness, which is 2.6mm, so stepping back from 2mm is also reasonable. 7T has much more CNR available, so pushing to minimum cortical thickness at 1.6mm makes sense there.

Temporal ICA is for removing global respiratory artifact. It is available in the HCP Pipelines GitHub, and we have been working on getting the automated classifier hosted as well.

Wishart Filtering is a better approach to reduce random thermal noise (which is what it sounds like you are talking about). How to use it best is a little complicated, though, and we plan to put a paper out on that. That paper is pending having some NHP connectivity data comparisons that will produce strong evidence for the benefits though.

Overall, the main potentially available technical advances would be a head only scanner capable of much shorter echo spacing (makes multi-echo viable) and some of the TIDY fMRI stuff where k-space info is shared across timepoints (amongst other benefits). I don’t have any TIDY fMRI data yet myself though, so I cannot comment on any limitations that technology might have.

Matt.

--
You received this message because you are subscribed to the Google Groups "HCP-Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hcp-users+...@humanconnectome.org.
To view this discussion visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/c1863699-2003-40f3-9283-9fbe614c4f05n%40humanconnectome.org.

The materials in this message are private and may contain Protected Healthcare Information or other information of a sensitive nature. If you are not the intended recipient, be advised that any unauthorized use, disclosure, copying or the taking of any action in reliance on the contents of this information is strictly prohibited. If you have received this email in error, please immediately notify the sender via telephone or return mail.

Michael Stevens

unread,

Mar 11, 2026, 7:16:08 PMMar 11

to HCP-Users, Glasser, Matthew

Thanks, Matt — incredibly helpful as always.

A quick follow‑up: Do you have guidance (here or offline) on the best way to implement Wishart filtering for HCP time series? I dug into it today, and I think I’ve pieced together the typical workflow:

concatenate the “clean” output from DedriftandResample,
load the dense data into MATLAB and de‑mean,
run SVD,
apply the Wishart filter, and
reconstruct the denoised data.

Even if that general flow is correct, having the steps isn’t the same as knowing I’m doing it right — e.g., selecting thresholds, handling surface vs. more problematic subcortical/cerebellar/brainstem voxelwise data, etc. And since I’m not using any HCP code for this yet (I assembled Matlab code from scratch for everything), there’s plenty of room for error. Could you offer some technical guidance on how you recommend implementing this? And the challenges you alluded to in your prior post?

We’re trying to see how much thermal noise-related problem we can mitigate across several datasets collected over the past 5+ years. We’re also choosing a sequence for new projects, including one starting in two weeks. So implementing Wishart filtering soon could help inform that decision.

Thanks again for your help,
Mike

Glasser, Matthew

unread,

Mar 11, 2026, 7:23:48 PMMar 11

to Michael Stevens, HCP-Users

Hi Mike,

It is a bit complex because some analyses benefit and others don’t and when you apply it also matters. I would prefer not to unleash the technique onto the world without the detailed guidance. Sorry for the delay on that. We have a mostly drafted manuscript, but I think it really needs the comparison to gold-standard connectivity to convince the skeptical reviewer (we have done it we preliminary data before, but we were hoping to get a final dataset together for this).

I would recommend the sequences below regardless of Wishart Filtering. They represent our best current approach. If you need it, I could provide a .exar1 (assuming you are on XA60 that is).

Also, Wishart Filtering is mainly important for dense analyses. Parcellation remains the best approach for reducing unstructured noise if you care about the effects at the level of brain areas (and Wishart Filtering doesn’t actually make much of a difference in the context of parcellated analyses because the parcellation effectively kills the unstructured noise already).

Michael Stevens

unread,

Mar 12, 2026, 4:25:06 PMMar 12

to HCP-Users, Glasser, Matthew, Michael Stevens

Thanks Matt... Again, incredibly helpful. I spent a bit more time yesterday looking into what folks seem to know about Wishart filtering and can begin to appreciate the complexities. No problem waiting until the approach is ready for release. I'll be on the lookout for a preprint; or keep me in mind when it's finally accepted?

And yes, if you could share a sequence we can load up into our Skyra, that'd be great. We're still running VE11C.2024 release however. Happy to grab it at an FTP site or any Github repository that's easiest.

Mike

Glasser, Matthew

unread,

Mar 12, 2026, 4:38:31 PMMar 12

to Michael Stevens, HCP-Users

I can’t build things for old software, just what I have on our MRPM server. I could send you a PDF if you want.

Harms, Michael

unread,

Mar 12, 2026, 5:47:00 PMMar 12

to hcp-...@humanconnectome.org, Michael Stevens

If you want, we have an E11C protocol for a Skyra for the AMP SCZ program that you could import as a starting point:

https://zenodo.org/records/14530974

cheers,

-MH

--

Michael Harms, Ph.D.

-----------------------------------------------------------

Professor of Psychiatry

Washington University School of Medicine

Department of Psychiatry, Box 8134

660 South Euclid Ave.

St. Louis, MO 63110

To view this discussion visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/7591EF2B-0D78-439E-A024-FE3A5138F906%40wustl.edu.

Glasser, Matthew

unread,

Mar 12, 2026, 6:08:13 PMMar 12

to hcp-...@humanconnectome.org, Michael Stevens

Is that what I suggested below or a different protocol?

Matt.

To view this discussion visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/2D2E3241-AC9C-433D-9CE5-774F7144E727%40wustl.edu.

Harms, Michael

unread,

Mar 12, 2026, 6:37:00 PMMar 12

to hcp-...@humanconnectome.org, Michael Stevens

It's 2.4 mm isotropic for the BOLD, with 60 slices, and TR=900 ms. Easy from there to modify the TR and MB factor (if allowed, for the chosen TR), if Michael wants to make changes

To view this discussion visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/D2F5BDE8-9411-48C3-831F-AC339EC3D505%40wustl.edu.

Glasser, Matthew

unread,

Mar 12, 2026, 6:39:24 PMMar 12

to hcp-...@humanconnectome.org, Michael Stevens

Skyra will require some fiddling to get that down. I can send a PDF later with all that already done.

To view this discussion visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/47350E15-4DCB-4D51-8FF3-4EDE2C636604%40wustl.edu.

Harms, Michael

unread,

Mar 12, 2026, 6:43:04 PMMar 12

to hcp-...@humanconnectome.org, Michael Stevens

The link I sent already contains a working Skyra protocol on E11C. The only question in my mind is whether one could lower MB to 5 at TR = 1000 ms.

To view this discussion visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/6A1DFA90-A243-494E-BA69-5E2F0E295F1A%40wustl.edu.

Glasser, Matthew

unread,

Mar 12, 2026, 10:47:52 PMMar 12

to hcp-...@humanconnectome.org, Michael Stevens

Sounds good. XA may well have broken Skyras anyway for fast EPI…

To view this discussion visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/AE1A4A97-D840-4886-86E9-AA94BF982919%40wustl.edu.

Michael Stevens

unread,

Mar 23, 2026, 9:34:31 AM (13 days ago) Mar 23

to HCP-Users, Glasser, Matthew, Michael Stevens

Thanks guys... this is incredibly helpful. I've been looking into this very hard this past week and I have a few follow-ups.

First, your intuition was right -- On our less-capable Skyra platform, I can get 2.4 mm voxels at TR=1 sec working at MB 5. We also tested MB 4 but need a TR of 1.20. Simulation/equation-based testing suggest both improve tSNR in most subcortical and some deep regions to not great... but perhaps workable levels (i.e., > 50 for MB 5, with a ~ 10 point tSNR advantage for MB 4 and the slower TR). But the MB 4 TR raises some concerns about cardiac alisasing that I'm wondering if you think can be mitigated with the standard HCP processing workflow? Some vague concerns if TR 1.2 is sub-optimal for HCP in general. But I thought you (and the listserv) might want to know what was achievable to be confident giving the same advice to others. We'll also test regional tSNR tests on actual BOLD data to make sure this is right.

Second, I've turned my attention to mitigating the problems for the gobs of data we've collected over the past 5 years on dozens of projects using that MB 7 sequence. Looking into what might be missing from our data processing, I realize I haven't yet applied tICA or created a sequence-specific ICA-FIX classifier set. So while I'm going to do that this week and reprocess, those don't affect the g-factor and thermal noise directly as I understand things, correct? So this leads to a few questions:

1. Matt, is the Wishart filtering approach available in a preprint form yet (or privately) so I can understand better how you're tackling the Wishart filtering approach to thermal noise mitigation? I'm going to sink some time into this... but with my luck, I'll devote a month to it only to see y'all release a far better approach once the paper is approved and code integrated into a new pipeline version. More to the point, I want to focus on testing clean-up options OTHER than what you're doing... with an eye towards compatibility. (Tangentially... last week I experimented with some success at adapting some Wishart approaches a stand-alone, post ICA-FIX step, with a 10-20 point tSNR gains in subcortex/brainstem without obvious artifact introduction to the entire CIFTI dataset... This likely will never been a widespread use tool, but perhaps useful for our sequence-specific poor signal quality needs).

2. Have you yet considered the downstream implications of magnitude-only NORDIC and/or MARSS (applied singly or sequentially) to the HCP preprocessing approach before the data enter the HCP pipeline? If you've already thought carefully about this and discarded those because of clear incompatibilities with HCP workflow, I'd like to know what you concluded the main issues are. I don't want to waste time testing a non-starter of an idea.

3. We've not yet switched to HCP v5.0 -- Is this the most stable release for automatic tICA implementation? We might as well clean the remaining physiological noise up optimally too while I work to improve overall subcortical/brainstem data quality in general. If I wanted a turnkey, bare-metal installation of an HCP pipeline version that included the best tICA denoising step implementation, what's the best option and source? Note, we've not adopted QuNex yet, in part because of uncertainty how strongly that platform will be developed or supported following Alan's move to industry (I'm aware he was a prime mover, but not aware of who might have committed to taking the lead on further work).

Finally, I'd welcome any other specific or general suggestions for things that I can try to optimize the use of the MB 7 deficiencies. I certainly don't mind looking into paths that the field hasn't specifically tried with HCP workflows.

Thanks,

Mike

Harms, Michael

unread,

Mar 23, 2026, 12:30:51 PM (13 days ago) Mar 23

to hcp-...@humanconnectome.org, Glasser, Matthew

Hi,

Since you've mentioned tSNR as a metric several times, I just want to make sure that you (and any others that see this) are aware of the Triantafyllou et al. series of papers:

https://www.ncbi.nlm.nih.gov/pubmed/15862224

https://www.ncbi.nlm.nih.gov/pubmed/16815038

https://www.ncbi.nlm.nih.gov/pubmed/21167946

Conceptually, one wants to maintain a "balance" between thermal and 'physiological ' noise, so the optimization problem is more complex than just looking at tSNR. (And separate from those papers, tSNR is also "incomplete" as an evaluation metric because it will increase with the longer TR due to the increased T1 recovery, but that doesn't factor in the decreased number of overall volumes that you get due to the increasing TR).

Matt can comment, but I'm not sure what all might have changed in the 'master' branch regarding the tICA-cleanup relative the latest tagged release of v5.0.0. Probably best to start with the 'master' branch if you want to do some tICA testing. [After we wrap-up some current testing/refinements of longitudinal processing code, we will probably tag a new versioned release].

Cheers,

-MH

--

Michael Harms, Ph.D.

-----------------------------------------------------------

Professor of Psychiatry

Washington University School of Medicine

Department of Psychiatry, Box 8134

660 South Euclid Ave.

St. Louis, MO 63110

To view this discussion visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/952829aa-3dbc-432b-8eaa-1f796f028882n%40humanconnectome.org.

Glasser, Matthew

unread,

Mar 23, 2026, 10:37:45 PM (13 days ago) Mar 23

to Harms, Michael, hcp-...@humanconnectome.org

I would aim for TR=1s.

Structured artifact cleanup is a good idea, but doesn’t directly relate to managing thermal noise.

This isn’t really an on-list topic right now, but off-list you can let me know what your applications are.
NORDIC in our hands created structured artifacts and reduced our ability to see neural signals. I had the opportunity to be a coauthor of MARSS, but did not believe it was a necessary step after working with the authors to produce some tests with their method.
We have not yet uploaded the automated tICA classifier, but this is planned soon. Right now, the HCP Pipelines are in a bit of a state of flux due to merging in the NHP streams. Hopefully, that will be completed in the next month or two. QuNex is doing just fine with Jure Demsar and Grega Repovs leading it.

I would recommend parcellated analyses unless you truly need dense analyses (which is often a minority of use cases).

With regard to Mike’s point about Triantafyllou et al., I am not really sure what to make of that work in the context of structured artifact removal. Ideally one would like artifact and noise free data. We have good methods to remove artifacts. We also have methods to remove noise, some of which are proven (parcellation) and some that are still under development (Wishart Filtering). If the neural signal is below the noise floor, it doesn’t matter what you do. If it is above, you can see it with these methods. His point about tSNR is correct though. We measure it in a modified way that excludes structured artifacts and we also do not penalize tSNR for neural signal variance. We also measure the ratio of neural signal variance to unstructured noise variance (fCNR).

Matt.

--
You received this message because you are subscribed to the Google Groups "HCP-Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hcp-users+...@humanconnectome.org.
To view this discussion visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/952829aa-3dbc-432b-8eaa-1f796f028882n%40humanconnectome.org.

Grega Repovš

unread,

Mar 24, 2026, 9:12:47 AM (12 days ago) Mar 24

to HCP-Users, Michael Stevens, Glasser, Matthew, Michael Stevens

Hi Mike,

Regarding:

> Note, we've not adopted QuNex yet, in part because of uncertainty how strongly that platform will be developed or supported following Alan's move to industry (I'm aware he was a prime mover, but not aware of who might have committed to taking the lead on further work)

Thanks for raising this—it’s a very reasonable concern.

Let me reassure you that QuNex is under active and ongoing development, with a committed team and a clear roadmap for the future.

It is true that much of QuNex’s functionality originated from the collaboration and the research needs of the Mind & Brain Lab in Ljubljana and the Anticevic Lab at Yale. Indeed, Alan played a key role in making QuNex publicly available, containerizing it, and implementing the initial turnkey functionality. That said, the core processing and analytical framework has been developed and is actively maintained by our team at the Mind & Brain Lab (myself and Jure Demšar). We also continue to collaborate with colleagues at Yale (the Cho lab).

We are also working closely with the HCP team to ensure that QuNex remains in lockstep with updates to the HCP Pipelines, and we are continuously improving the platform in terms of robustness, flexibility, and usability. The evolution from the original turnkey workflows to the more flexible run_recipe framework is one example of ongoing development and improvements to QuNex (see: https://www.biorxiv.org/content/10.1101/2025.11.08.687330v4). We have several further improvements in the pipeline aimed at making extensions easier to develop, improving codebase flexibility, simplifying logging and others. Development is continuing at full pace.

Importantly, we “eat our own dog food”: QuNex is a core part of our own research toolbox, we rely on it extensively for neuroimaging analyses, which ensures that development is driven by real-world use and that issues are addressed quickly. We are responsive to the user community via the forum (https://forum.qunex.yale.edu) and are happy to incorporate feedback into development.

In short, QuNex is very much alive, actively used and maintained, and has a clear path for continued development.

All the best,

Grega

Stevens, Michael

unread,

Mar 24, 2026, 9:28:50 AM (12 days ago) Mar 24

to hcp-...@humanconnectome.org, Michael Stevens, Glasser, Matthew

LOL, I love the way you put that.

Quick question – I’m about to jump through hoops to make a bare metal install of HCP v5.0 work on our specific linux infrastructure… Most notably struggling with the complex binary library incompatibilities of FreeSurfer 6.0 and our Debian 12.12 OS. We maybe could swipe all that away by moving to QuNex for v5.0. My specific hesitation is that my read of available guidance is that the tICA in the current QuNex container doesn’t yet include all the automatic, single-subject application updates that Matt and others have worked to implement. Do I have that wrong? If QuNex offers all the functionality currently in the MASTER repo (obviously not the Wishart filtering Matt’s still developing), it probably would be worth it to switch over immediately and bear the learning curve. A second question is to make sure I know what container environment… We currently only have singularity up and running on our newer local OS architecture I’m hoping to use moving forward.

Please let me know if you can clarify, with my thanks!

Best,

Mike

From: Grega Repovš <grega....@gmail.com>

Sent: Tuesday, March 24, 2026 9:13 AM
To: HCP-Users <hcp-...@humanconnectome.org>

Cc: Michael Stevens <mcste...@gmail.com>; Glasser, Matthew <glas...@wustl.edu>; Michael Stevens <hcp-...@humanconnectome.org>
Subject: Re: [hcp-users] High frequency noise related to MB and temporal ICA

Hi Mike, Regarding: > Note, we've not adopted QuNex yet, in part because of uncertainty how strongly that platform will be developed or supported following Alan's move to industry (I'm aware he was a prime mover, but not aware of who might

To view this discussion visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/d47214a9-b756-46cd-a0c8-ac78d0876019n%40humanconnectome.org.

This e-mail message, including any attachments, is for the sole use of the intended recipient(s) and may contain confidential and privileged information. Any unauthorized review, use, disclosure, or distribution is prohibited. If you are not the intended recipient, or an employee or agent responsible for delivering the message to the intended recipient, please contact the sender by reply e-mail and destroy all copies of the original message, including any attachments.

Glasser, Matthew

unread,

Mar 24, 2026, 9:30:29 AM (12 days ago) Mar 24

to Stevens, Michael, hcp-...@humanconnectome.org, Michael Stevens

We haven’t made automated tICA classification available yet to anyone. The current issue has been hosting the model (which is a bit large for GitHub), but we will be working through that.

Matt.

--
You received this message because you are subscribed to the Google Groups "HCP-Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hcp-users+...@humanconnectome.org.
To view this discussion visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/d47214a9-b756-46cd-a0c8-ac78d0876019n%40humanconnectome.org.

This e-mail message, including any attachments, is for the sole use of the intended recipient(s) and may contain confidential and privileged information. Any unauthorized review, use, disclosure, or distribution is prohibited. If you are not the intended recipient, or an employee or agent responsible for delivering the message to the intended recipient, please contact the sender by reply e-mail and destroy all copies of the original message, including any attachments.

Harms, Michael

unread,

Mar 24, 2026, 2:14:37 PM (12 days ago) Mar 24

to hcp-...@humanconnectome.org, Michael Stevens, Glasser, Matthew

In general, QuNex is built off the current 'master' branch of HCPpipelines at the time of the QuNex build. We don't limit QuNex to use only tagged HCPpipelines versions.

And yes, there is a Singularity/Apptainer version of the QuNex container (which is what we use all the time at WashU).

Cheers,

-MH

--

Michael Harms, Ph.D.

-----------------------------------------------------------

Professor of Psychiatry

Washington University School of Medicine

Department of Psychiatry, Box 8134

660 South Euclid Ave.

St. Louis, MO 63110

From: "Stevens, Michael" <Michael...@hhchealth.org>

Reply-To: "hcp-...@humanconnectome.org" <hcp-...@humanconnectome.org>
Date: Tuesday, March 24, 2026 at 8:29 AM
To: "hcp-...@humanconnectome.org" <hcp-...@humanconnectome.org>
Cc: Michael Stevens <mcste...@gmail.com>, "Glasser, Matthew" <glas...@wustl.edu>

To view this discussion visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/PH0PR16MB4214ECD23C56641E64463C81F448A%40PH0PR16MB4214.namprd16.prod.outlook.com.

Reply all

Reply to author

Forward