Attempting to process mouse brain functional ultrasound data with RABIES fails if more than one subject

[benhibou]

Describe the bug
We are attempting to use RABIES to process functional ultrasound data (recorded in mouse brain) as we read it should be possible. We are trying to run RABIES with the flag"--bold_only". If a single individual is submitted for processing we are able to go pretty far in the processing with RABIES, in fact all the way to getting a connectivity matrix in the analysis step. As soon however as we submit for processing more than one individual, the RABIES processing crashes already in the preprocess step.

Describe RABIES call
Include a copy of the command call you executed from the terminal, and any additional information that could be relevant to your execution:
For preprocess:
/nexus/posix0/MPI-psych/g/CU/MR/bebou/mon_rabies/rabies.sif -p MultiProc preprocess /nexus/posix0/MPI-psych/g/CU/FUS/BB/DSURQE/test/r3/ID00071/r3_ID00071_qui_two /nexus/posix0/MPI-psych/g/CU/FUS/BB/DSURQE/test/r3/ID00071/r3_ID00071_qui_prep_two_nr --bold_only --anat_template /nexus/posix0/MPI-psych/g/CU/FUS/BB/DSURQE/test/r3/ID00071/r3template/r3rDSURQE_40micron_average.nii --brain_mask /nexus/posix0/MPI-psych/g/CU/FUS/BB/DSURQE/test/r3/ID00071/r3template/r3rDSURQE_40micron_mask.nii --WM_mask /nexus/posix0/MPI-psych/g/CU/FUS/BB/DSURQE/test/r3/ID00071/r3template/r3rDSURQE_40micron_WM_mask.nii --CSF_mask /nexus/posix0/MPI-psych/g/CU/FUS/BB/DSURQE/test/r3/ID00071/r3template/r3rDSURQE_40micron_CSF_mask.nii --vascular_mask /nexus/posix0/MPI-psych/g/CU/FUS/BB/DSURQE/test/r3/ID00071/r3template/r3rvascular_mask.nii --labels /nexus/posix0/MPI-psych/g/CU/FUS/BB/DSURQE/test/r3/ID00071/r3template/r3rDSURQE_40micron_labels.nii --apply_STC --commonspace_reg masking=false,brain_extraction=false,template_registration=SyN,fast_commonspace=false

rabies_preprocess.log

Attach log file
Attach to your issue the .log files present in the output folder. (e.g. rabies_out/rabies_preprocess.log)
Attached

Attach QC_report
Attach to your issue the QC_report folder present in the output folder. If QC_report is too large, mention that it couldn't be shared, and if possible, provide an alternative access to the files.
Attached

Additional context
Add any other context about the problem here.

[gdevenyi]

Hi.

Functional ultrasound processing is entirely experimental. There are in-process details at #445

I will direct you to @milaurose who is spearheading this work for help/comments.

[Gab-D-G]

Hello @benhibou nice to meet, really glad to see you're trying to make RABIES work with fUS! It's definitely more of a niche application, but a few people have succesfully achieved it in the past, and as Gabe mentioned, we are currently implementing some new options that are specifically designed to support that kind of data.

These updates should be available in the coming days as a new RABIES version, so stay tuned. However I refer to @milaurose for more specific implementation details.

One comment I can make for now though is that it definitely is not ideal to use an MRI template as a reference for registration. If you could provide instead some form of angiogram template that better ressemble your fUSI data, I believe that would enhance the processing greatly.

I am surprised actually that it worked on a individual subject level... but the error you get when using multiple subject is better there is some kind of massive registration error in the template generation step - this would probably be fixed with a better reference template.

[milaurose]

Hi! So sorry for the delayed response - we have finally merged the modifications to enable processing fUS scans in RABIES into the main branch. The new options will be available soon in the next rabies version. This functionality can still be improved, but at least you should be able to get decent outputs for all your scans! Below I'm including the preprocessing call that works for my data (note, my data was acquired with the Iconeus MultiArray probe, 16 slices with a TR of 2.4s, and I also have an angio scan from the same session with 60 slices and 10 averages - if you have fewer than 16 slices and/or they are spaced far apart, then the registration may get trickier...)

Sample preprocessing call:

rabies -p MultiProc --local_threads 15 preprocess $bids_folder $out_folder \
--log_transform \
--anat_inho_cor method=disable --bold_inho_cor method=disable \
--commonspace_reg masking=false,brain_extraction=false,keep_mask_after_extract=false,template_registration=Affine,winsorize_lower_bound=0.005,winsorize_upper_bound=0.995,stages=rigid-affine \
--bold2anat_coreg masking=false,brain_extraction=false,keep_mask_after_extract=false,registration=Rigid,winsorize_lower_bound=0.005,winsorize_upper_bound=0.995 \
--anat_template $template --brain_mask $mask \
--WM_mask $wmcsf  --CSF_mask $wmcsf  --vascular_mask $mask  --labels $labels

Explanation:

• As Gabe mentioned, you should provide a template that resembles the input images, like a high-quality angio scan. The MRI template that is the default in rabies doesn't have the same vascular contrast as the fUS scans, so it won't be able to align vasculature across scans.
• the log_transform option will compute the log of the scans and use this for the registration steps (your final output will still have the original intensities). We have found that this improves the registration accuracy. In order for this to work properly, the template that you provide also needs to be log transformed (rabies won't do that part, you need to compute it yourself and provide that path).
• disable inhomogeneity correction because this option exists primarily to correct for B1 inhomogeneity in MRI scans (less relevant here). fUS has its own inhomogeneities, but it needs to be validated whether this technique can deal with them.
  use only Affine for registration to the template and only rigid for registration between the functional and angio scans. Non-linear registration could work but it's better to start simple and switch to non-linear only when you're confident that affine and rigid work. Personally, I'm not so confident with using non-linear registration given the vascular nature of the fUS contrast - I'm not sure that there is enough information about brain shape to make precise local deformations.
• I do really recommend not going with the --bold_only and instead actually providing an angio scan in place of the MRI anatomical. The high-SNR and number of slices of the angios really helps with the registration, and plus they're so fast to acquire that there's really no reason to skip them.
• STC can be tricky depending on the type of probe you have (for example, the array probe is acquiring 4 slices simultaneously). You can start without any STC then refine it after the pipeline is working for you. I personally have not gotten around yet to figuring out how to best apply STC to my fUS data.

Other tips

• As I implied above, these recommendations will depend on the fUS sequence that you're using. If your number of slices, probe type, slice spacing, FOV or timing is very different from what I mentioned, then please follow up with your parameters so that we can give a more tailored recommendation.
• If you want to join the fUS community to ask for tips (or maybe a template), join us on discord: https://discord.gg/V6gAnb5bHr (this link will only work once).