IQMs values differ between two MRIQC containers.

[Hasnae Ag]

Dear all,
I'm using MRIQC 22.0.0rc1 and I've created a new one called 'local mriqc' based on that, with the exception of replacing the secret token in the mriqc folder.
I noticed variances in various IQMs for the T1 modality after running both containers on the same dataset. "cnr", "qi 2", "summary bg k", "summary bg mean", "summary bg n", "summary bg stdv", "tpm overlap csf", "tpm overlap gm", "tpm overlap wm", "tpm overlap csf", "tpm overlap gm", "tpm overlap".


I also tried running the 'local mriqc' twice on the same dataset, however, the same IQMs for T1 modality were different each time. The same issue happens with 'nipreps/mriqc: latest'. 

Has anyone ever attempted a comparison like this? I've included several screenshots as examples. Note that the dataset is valid BIDS. 


Any assistance would be much appreciated!

Hasnae 

[Oscar Esteban]

Hi Hasnae,

Thanks for your question. This is not surprising -- some of the masks used to calculate those metrics are derived from an image registration process implemented with ANTs. To ensure run-to-run reproducibility to the last bit, you would need to (i) ensure you set the same random seeds for all processes affected, ANTs and Python, mostly; (ii) run ANTs in single-thread mode at the cost of speed down.

For the purpose of MRIQC we considered that some variance is acceptable and didn't put much thought into this (conversely, with fMRIPrep you have fine control of the seeds, for instance).

HTH

Cheers,

oe

--
You received this message because you are subscribed to the Google Groups "mriqc-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mriqc-users...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/mriqc-users/e1b307eb-81d1-45e5-a831-6fcdf07604ffn%40googlegroups.com.

[Hasnae Ag]

Hi, 

Thanks for the quick answer and explanations. 

Hasnae  

To view this discussion on the web visit https://groups.google.com/d/msgid/mriqc-users/CAAsFcgwtjcUrsjKHW1e%3DqdjGpZk6Rn%2BF7Tu2n%2BPbzmtZsa26gA%40mail.gmail.com.