QSDR Registration Issue with Ex-vivo Mouse Brain Data (Low R^2 Score)

[Zhuoheng]

Dear Frank,

I am working on ex-vivo mouse brain dMRI data processing, specifically using QSDR in DSI Studio to generate template space *.fib.gz files.

I've consistently run into an issue with the registration accuracy, which I believe is affecting the final quality of the mapped data.

Problem:
Low R-squared Score: The registration R-squared score is consistently low, hovering around 70% to 80%.
Artifacts on dti_fa: The dti_fa map (as shown in the attached image) clearly displays unusual, high-intensity pixels at the periphery of the brain. This suggests that the normalization/registration is not accurately mapping the boundaries, even though my brain mask has been smoothed (mask.gz file). QA image does not have this issue.

Steps I've Taken:
I've ensured the initial brain mask is clean and smoothed.
I attempted using a custom template that I created using ANTs from my own data set, but the registration issue (low R-squared and peripheral artifacts) persists.

Questions:
I suspect the registration process itself is the bottleneck for achieving a more accurate template space result.

How can I improve the R-squared score and the overall registration accuracy, particularly for ex-vivo mouse data?
Are there any additional parameters within DSI Studio's QSDR implementation (or the underlying registration tool) that I can adjust, similar to the extensive transformation/metric/iteration controls available in toolkits like ANTs (e.g., rigid, affine, SyN stages, multi-resolution levels, different similarity metrics)?

Any guidance on optimizing the registration for this specific type of data would be greatly appreciated.

Thank you very much for your time and help.

Best regards,

Hopper

[Frank Yeh]

DWI is known for low SNR, and thus low R2 can be common.
It should be fine if anatomical landmarks aligned well with the atlas.

Moreover, the absolute value of R2 itself cannot assert whether the alignment is good/bad.
The value should be evaluated within a group.
I usually collect a group distribution of R2 to identify outliers (e.g. volume flipped upside down, or other issue) for correcting or dropping problematic scans.

--
You received this message because you are subscribed to the Google Groups "DSI Studio" group.
To unsubscribe from this group and stop receiving emails from it, send an email to dsi-studio+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/dsi-studio/88ae2f45-17b5-44dd-94fb-6907193ac11cn%40googlegroups.com.

[Hopper]

Hi Frank,

Thank you for the quick reply. 

I have a couple of questions focused on registration issues for my ex-vivo mouse brain data.

During the QSDR process, are there any parameters that I can set or tune to directly improve the registration accuracy?

I also find some edge artifacts. Why are there extra voxels (a 'halo') appearing on the dti_fa map at the brain's edge? I suspect this artifact is severely impacting my registration results. What should I do to remove or prevent this edge artifact? Should I re-mask the resulting data in the template space, or is there a specific setting in QSDR I can adjust?

Thank you very much for your advice.

Best,

Hopper

[Frank Yeh]

I also find some edge artifacts. Why are there extra voxels (a 'halo') appearing on the dti_fa map at the brain's edge?

This is due to low DWI signals that lead to ill-posed estimation of the tensor model which corrupts FA.
 

I suspect this artifact is severely impacting my registration results.

DSI Studio does not use FA in registration. It uses QA. 
 

What should I do to remove or prevent this edge artifact?

Not much you can do because regardless of any cutoff point, the uncorrupted FA value will blend with meaningful FA value. A rigid cutoff will create a stronger edge, which could be even more problematic.
 

Should I re-mask the resulting data in the template space, or is there a specific setting in QSDR I can adjust?

There are indeed parameters in the nonlinear and linear registration, and I will see what I can do to allow customization.

Best regards,
Frank
 

To view this discussion visit https://groups.google.com/d/msgid/dsi-studio/5ad36703-d485-4122-89b2-cef4d64090a8n%40googlegroups.com.

[Frank Yeh]

There are indeed parameters in the nonlinear and linear registration, and I will see what I can do to allow customization.

DSI Studio is now updated to provide the following parameters with QSDR command line:

dsio_studio --action=rec --source=some.sz --method=7 --reg_speed=0.3 --reg_smoothing=0.05

Higher 'speed' handles large distortions, and high smoothing ensures more structure continuity.