Optimal combination of multi-echo ASL

Hello folks,

We are running an ASL/BOLD multi-echo multiband sequence over a TR of 4.115 seconds.
We would like to extract the ASL signal with the best tSNR from the BOLD signal (label vs control).

In the litterature, the usualy filter and demodulate one echo, but we were wondering if it would be possible to use the 4 echos of the label and echos to gather as much signal as possible.
Something like tedana would be marvelous.

This sounds like a good idea, but I’m not sure if anyone has specifically tried it yet. I don’t keep up with the ASL literature, but I doubt there is public software that could do it for you. As a first pass, I’d lean away from ICA-based denoising because the out-of-the box methods of selecting good vs bad components likely won’t work for the tagged volumes.

You probably could do the weighted average of the echoes (i.e. Poser et al 2006 https://doi.org/10.1002/mrm.20900 ) but you might want to give some thought about what weightings optimize for what your want in the tag & control image. Any consistent weighting would likely be better than not using all 4 echoes.

Hope this helps


Hello, thank you for sharing your thoughts.
We are trying ICA denoising, but tedana doesn’t converge once the echos are filtered. I’ll have to check on this.

Regarding the optimal combination of ASL, we were tinking of giving same weights on label and control.
The best option would be to fit an exponentiel curve I guess.

As a start, what do you think of simply averaging the filtered/substracted echos.

It would look like
1/4(echo1LBL-echo1CTRL + echo2LBL-echo2CTRL +…+ echoXLBL-echoXCTRL).

It is a broad approximation, but a start.
One of the question we are wondering is on which volumes to mean the echos.
Should we mean on raw ASl, or filtered volumes (CBF), or once the echos have been regressed ?

Thanks for any advice.

I’m thinking about this a bit more.
My understanding is, for single-echo ASL, the minimum possible TE is used because T2/T2* decay is an artifact. That means averaging the echoes, as you show would improve the overall SNR at the cost of having a bit more T2/T2* fluctuations that can potentially contaminate the results. I suspect you’d benefit from a weighted average that biases towards the shortest echo, but I haven’t thought about a way to choose optimal weights for ASL.

Also with this in mind, you definitely wouldn’t want to use the standard tedana ICA denoising because that is trying to maximize T2* (BOLD) fluctuations. That said, if you calculate your LBL-CTRL for each echo separately, there might be an interesting T2* variation that could be used for ASL-specific noise removal (figuring that out might be a stand-alone project).

I’m not completely sure what you mean by ‘which volumes to mean the echos’

1 Like

Hello, and thanks for carrying the topic.
I was thinking about it myself today.
When I talk about volumes, I refer to datasets, at chich step of preprocessing shall we mean.

Today we were wondering if it was good to use the T2 to optimise the combination, but you say it is an artifact. How so ?
As far as I understant, labelling would be marking with an angle the blood, won’t this label be afected by the T2 decay ?
Is there an optimal T2 to find for the blood to maximise the Ctr-Lablel, as it is done in tedana ?
We were gonna apply a maximisation of the difference (labeling of blood between the ctrl and lbl) for all 4 echos. If there is an optimum for a certain echo that can be reach for each voxel, we might keep this one echo, or ponderate it higher.

Wouldn’t this decay be the same for the control and the label, and thus not afecting the difference between the two ?

May you be a little more specific on the artifact part of the T2 ?


I’m not an ASL person so I’d recommend you also ask these questions to someone with a bit more expertise in this area than I have. There are many different ASL methods, but for several ASL methods, you are tagging everything in a slab and then seeing how many of those tagged protons end up somewhere else (i.e. because the protons are part of flowing blood). The spin labeling will decrease over time and T2/T2* is one factor that can affect how much the signal decreases. Since T2* is also fluctuating over time (i.e. the BOLD effect) that might interact with the tag-control measurements. The shorter the TE the less variation in T2* will be a factor.

Like I said, I’ve used ASL a couple of time, but I’m not really an ASL person so run this by someone who understands ASL better than me to confirm my explanation.

Hope this helps.

@ztn I just wanted to follow up on this. Have you figured out what to do with your ASL data?

We’ve gone classic ourselves.
Being a PhD candidate, I don’t have much time to produce such work.
I am looking for few people to collaborate on it though.

If you or anyone else wish to have a look at this, please contact me.

1 Like

We are having a conversation about it with some people of TEDANA team here.