Mixing two HARDI acquisition protocol in analyses

Hi all,
I have two cohorts of neonatal subjects.

    1. The 1st one acquired with 9 b=0, 45 b=1000 and 45 b= 2000 s/mm2
    1. The 1st one acquired with 6 b=0, 30 b=700 and 60 b= 2800 s/mm2
      Both groups include term and preterm born infants scanned at TEA, both healthy and pathological (encompassing a variety of anomalies).
      I want to analyze those subjects with advanced microstructural models (DKI, NODDI, MSMTCSD…) and perform structural connectomes.

My question is:
→ can I use the whole ensemble of subjects made up of two different acquisition protocols?
→ does this depend on the final/clinical aim of my analysis?
thank you

1 Like

Hi Rosella,

OK, there are two levels to this question:

  • Is it possible to process data acquired using different acquisition protocols to produce equivalent outputs – at least in principle?

  • Given the inevitability of protocol-induced differences in these outputs, how to account for these in the analysis?

The latter is a common issue in multi-site studies, and generally relies on including site/scanner/protocol as a covariate in your analysis. This can only work if you have balanced groups across sites/protocols.

The first issue is an interesting one. For MSMT-CSD, the issue is to find a way to obtain a group-average response function that works across both groups. While possible in theory, this would require quite a bit of development to implement, and isn’t something that I foresee being available any time soon. Furthermore, even if this were possible, I doubt it would remove all differences in the outputs… You could instead derive group-specific average responses and use them, followed by mtnormalise to try to mitigate any intensity scaling issues, but I would personally be concerned about the potential that the responses would be capturing subtly different aspects of the tissue – especially on a neonatal cohort. That said, if site/protocol is included as a covariate, and your groups are fully balanced, I expect these effects would largely be factored out, but I would nonetheless be cautious with that approach.

These issues would I expect also apply for other analysis methods, including DKI and NODDI – though they may manifest somewhat differently. NODDI is likely to be the least affected given the nature of the model, but I would still expect differences due to the choice of protocols. This may be something that has already been investigated elsewhere, but I haven’t come across such work myself. Others on this forum will hopefully have more helpful insights on that front…

Sorry if this isn’t all that helpful…
All the best,

1 Like

thank you very much !

the question from Rosella is very interesting. I would like to ask a question related to the analysis from different acquisitions. For example, I have two datasets of two different diseases, I call them disease 1 and disease 2. The DTI is very similar, but the data were acquired from different scanners, site, and different parameters. Therefore, disease 1 from acquisition 1 and disease 2 from acquisition 2. I would like to compare the two groups. I think that I cannot use site/scanner as covariate because this covariate already distinguishes the two groups. Also I do not think that it is a good idea to analyze the DTI metrics from two different DTI acquisitions. However, my question is: is it possible to analyze the connectivity matrices from two different acquisitions or, also, the connectivity has the same issue that the DTI-metrics have?

Hi Andrea,

I’m by no means on expert on multi-site studies such as this one, but my understanding is that it would be extremely difficult to do this kind of analysis with the data you have. As you can already see, including site as a covariate when the two groups are collected at different sites would immediately absorb any group-wise differences and kill any between-group effect. The connectivity information would be affected just as much as any other metric – or at least, I don’t think it would be reasonable or defensible to assume that the effect of scanner was negligible.

What you might be able to do though, is perform correlation analyses, where you would still be able to include scanner as a covariate – but that relies on you having data compatible with such an analysis, and there being a good spread of effect size across both groups.

The other study design that might work in this context is a longitudinal study (e.g. looking at the difference pre/post treatment), but I’m not convinced this would necessarily allow you to avoid including site as a covariate – which may then also absorb some or all of the effect you’re looking for. It’ll really depend on the details of the analysis and how convincingly you could demonstrate that the effect you’re seeing is not driven by scanner differences.

Hope this helps,

1 Like

Hi Donald,
yes, you helped me a lot.
I had already imagined this… :wink: