CRLBs vs LCModel

[Jonathan Underwood]

Hi Martin et al,

I've been analysing Siemen's SVS using Tarquin. Most of the data seems to be fit reasonably well but the %SDs seem to be quite high (sometimes over 100%) in some which makes me concerned that metabolite concentrations aren't accurate. We don't have a licence to LCModel at my insitution but via a helpful colleague we've run a few of the more dodgy spectra through LCModel to compare the %SDs etc. The manual for LCModel states that their %SD is half the CRLB so am I correct in saying that they are also half the %SDs in Tarquin (which I thought were the same as the CRLB)? Even so, the %SDs seem markedly lower in LCModel. From what I can see the baselines seem to be better modelled in Tarquin but the fits look okay for both on what are not perfect spectra.

I've attached a couple of examples (.png for LCModel and .pdf for Tarquin in case it isn't obvious).

My main question is does this mean the LCModel results are more accurate as they have lower %SDs or does this not really matter particularly when averaging over multiple subjects? Or are the spectra of such poor quality that they just shouldn't be used.

Also although MI, tNAA, tCho and tCr broadly agree Glx seems to be consistently half that found in LCModel. Is this something you've noticed before (I didn't see it mentioned in your paper comparing TARQUIN and LCModel)?

Thanks in advance,

Jonathan

[Martin Wilson]

Dear Jonathan,

Thanks for the comparison. CRLB estimates require a measure of the noise level and it is possible the noise is being overestimated in the time domain due to an artifact or incomplete water removal.

If you're able to send an example data file I can check. I'll answer your other questions after this.

Martin

--
You received this message because you are subscribed to the Google Groups "TARQUIN users' group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tarquin_users_g...@googlegroups.com.
To post to this group, send email to tarquin_u...@googlegroups.com.
Visit this group at http://groups.google.com/group/tarquin_users_group.
For more options, visit https://groups.google.com/d/optout.

[Jonathan Underwood]

Hi Martin,

Thanks for your swift reply.

I've emailed the raw data to you.

For some of them there's the issue of incomplete water suppression.

Will be interested to hear what you say.

Thanks,

Jonathan

[Martin Wilson]

Hi Jonathan,

I've had a quick look at your data and it looks like you have a clear
"spurious echo" (see attached). This signal is being considered as
part of the noise estimate by TARQUIN and explains the large
differences in the CRLB estimate compared with LCModel (which must get
the noise estimate from the frequency domain signal).

For a fix, I have been thinking about estimating the noise in the
frequency domain. However, I don't think the increased errors are
"wrong" as these artefacts will likely contribute to some additional
error.

For the Glx differences, it has been noted that LCModel has some bias
with increasing lineshape (TARQUIN may also - but this hasn't been
tested) so this may explain the result. However, reliable Glx measures
require good quality data - so I think poor data quality is the
overriding issue here.

For extra information on spurious echos, lineshape fitting bias and
general data quality I recommend a look at : "Issues of spectral
quality in clinical 1H-magnetic resonance spectroscopy and a gallery
of artifacts" NMR Biomed. 2004;17:361-381 by Roland Kreis. Figure 14
shows the "spurious echo" problem.

Martin

On 7 October 2014 16:46, Jonathan Underwood

[Jonathan Underwood]

Thanks Martin.

The spurious echo definitely looks to be an issue in the 0142 ACC but is this the same for the putamen spectra, which are have a lower SNR (presumably due to iron concentration etc)? Or it is to do with the different way of estimating noise.

When you compared TARQUIN with LCModel on a large sample of clinical data did you find the CRLBs were larger in TARQUIN because of the estimate of noise is based in the time domain vs the frequency domain? My slight concern is that when it comes to reporting results some may worry about the high SDs in comparison to previous work done with LCModel using a 30% cut off for QC and reporting.

In any case should I have as much faith in the concentrations compared to LCModel as the CRLB estimates are greater in TARQUIN due to the different way of estimating noise or do they represent greater uncertainty in the metabolite concentration?

With regard to the Glx number this seems to be the case for better quality data (see attached) too. Is this simply because my better quality data is simply not that great?

Sorry if these are elementary questions by the way.

Thanks again,

Jonathan

[Martin Wilson]

Hi Jonathan,

I can take a look at the example you provided, it's possible there is
a similar issue, but just not within the spectral window. Have you
sent me this spectrum before?

Martin

On 8 October 2014 12:21, Jonathan Underwood

[Jonathan Underwood]

Hi Martin,

I sent you the raw data for two subjects for the ACC and putamen I mentioned in my first post. I think the spurious echo and ghosting is an issue for the ACC spectra but the putamen ones less so.

I'll send you the raw data for the other one for the Glx question too.

I'll be interested to know what you think.

Thanks,

Jonathan

[Martin Wilson]

Dear Jonathan,

This data looks pretty decent, and it does seem that there are some
differences in the Glx concentration between TARQUIN and LCModel for
this case - most likely due to differences in the baseline estimate.

It has been noted previously (Roland again, and more recently Jamie
Near) that LCModel metabolite concentrations are inversely
proportional to linewidth, this may be a factor here. I wouldn't like
to say which measure is the most likely (I'm biased anyway!), however
I wouldn't be too concerned about CRLBs. The most important thing is
to convince yourself there is a peak there, between 2.2 and 2.4 ppm I
see something that looks very much like glutamate so I would include
this result in a group analysis regardless of CRLB.

I intentionally calculate the CRLBs based on the assumption that the
baseline is poorly determined, it's possible that LCModel assumes the
baseline is well known - and this would explain the difference.

Hope that helps,

Martin

On 9 October 2014 10:32, Jonathan Underwood