Questions about general pipeline

[nicholasjchua]

Hello! I am using model_build to create a template for whole salamander brains imaged with a lightsheet microscope. I ran an initial test with 6 samples using mostly default parameters, e.g., iterations = 4, stages = (rigid, similarity, affine, nlin). The last 2 nlin iterations haven't finished yet, but I have been looking at the (extremely promising) results and have some questions about how the outputs of each stage/iteration interact with each other.

(if there is a paper or flowchart describing the general approach, it may answer all my questions)

1. I noticed that 4 iterations of each stage are computed before moving on to the next type of registration. Is this correct? I am used to pipelines were each iteration involves computing affine and SyN for each dataset, averaging the transformed data, and then updating the template for the next iteration. Is optimized_antsMultivariateTemplateConstruction doing something different?

2. What is happening in the similarity stage? This is not a 'type' of ANTs registration I am familiar with.

3. If I don't specify shrink_factors, are all the transformations computed on the full resolution data?

Thank you for creating this!

[gdevenyi]

Hi,

I noticed that 4 iterations of each stage are computed before moving on to the next type of registration. Is this correct? I am used to pipelines were each iteration involves computing affine and SyN for each dataset, averaging the transformed data, and then updating the template for the next iteration. Is optimized_antsMultivariateTemplateConstruction doing something different?

The conceptual steps this pipeline takes are the same as antsMultivariateTemplateConstruction2.sh.

Best way to thing of the default here is this is equivalent to running antsMultivariateTemplateConstruction2.sh 4 times, with a rigid, then a similariy, then an affine, then an nlin, and passing forward the average from each as the initialization of the next.

Note that the nlin stage is the "regular" way that antsMultivariateTemplateConstruction2.sh runs, and includes a full rigid+similarity+affine+nlin at each iteration. Alternatively, you could run nlin-only.

I have tried to document the similarities/differences here, please let me know if you want more detail anywhere:
https://github.com/CoBrALab/optimized_antsMultivariateTemplateConstruction#differences-from-antsmultivariatetemplateconstruction2sh

What is happening in the similarity stage? This is not a 'type' of ANTs registration I am familiar with.

Similarity is Rigid+ 3 Scale, its available in ANTs but not used in their default tools anywhere.

If I don't specify shrink_factors, are all the transformations computed on the full resolution data?

The registration iterations/shinks/smooths are computed automatically from the size/resolution of the fixed image using antsRegistration_affine_SyN.sh and ants_generate_iterations.py. Every iteration goes to the full resolution of the data, unless you set the "volgenmodel" mode, which is a different modelbuilding method which was originally used to generate the ICBM MRI models ~20 years ago.

[nicholasjchua]

Wow, thank you for your quick and extremely informative reply.

I will read through the ANTs multivariate pipeline you mentioned. Just to make sure I am understanding you correctly, if I set iterations to 4, is the average from the 3rd iteration of rigid used as the template for the 4th rigid? Subsequently, is the output from the 4th rigid iteration then used to initialize the 1st similarity registration, and so forth?

Thanks for clarifying the points about similarity and scale! I understand now.

[gdevenyi]

is the average from the 3rd iteration of rigid used as the template for the 4th rigid?

yes, as INITIALIZATION, the template evolves after with updates

is the output from the 4th rigid iteration then used to initialize the 1st similarity registration, and so forth?

yes, as INITIALIZATION, the template evolves after with updates

[niksirbi]

I made the following schematic while using model_build.sh to create a bird brain template (from serial 2-photon microscopy data).
Similarly to you @nicholasjchua I used the default 4 iterations and 4 transform stages, and as you can see there's definitely a progressive improvement in the template's "sharpness".

My schematic shows the order of execution, which @gdevenyi seems to confirm above. It also shows the allowed transform types per stage. I do have a clarifying question though:

Similarity is Rigid+ 3 Scale

The above answer suggests similarity = rigid + xyz scales, while my schematic (possibly falsely) indicates similarity = rigid + 1 uniform scale. Would you mind confirming which one is true @gdevenyi? (As always, thanks for this outstanding piece of engineering).

[nicholasjchua]

This is fantastic!!! The zig-zag arrow showing the order is exactly what I was wondering. Before running it, I assumed it would be rigid0 -> similarity0 -> ... -> rigid1, and so forth, but as @gdevenyi described, all iterations of a transform stage are completed before moving to the next type of transform.

[gdevenyi]

You're right, I'm mixing up MINC's lsq9 (rigid + 3 scales) with ITK Simlarity (rigid + 1 scale)

https://itk.org/Doxygen/html/classitk_1_1Similarity3DTransform.html

My bad.