Meaning of MI.1D produced by SUMA

Hello,

I want to quantify the accuracy of alignment for surfaces that are aligned to std.141 during the normal SUMA pipeline. In other words, how much error is there in the alignment, akin to a cost metric that is produced in traditional anatomical alignment.

I think way to evaluate alignment is to use the std.141.*.MI.1D files that are produced by @SUMA_Make_Spec_FS. I think that this file measures the mutual information cost of the alignment of each node to the std.141 mesh; however, I am not sure this is true. It seems to be created by MapIcosahedron (https://afni.nimh.nih.gov/pub/dist/doc/htmldoc/programs/MapIcosahedron_sphx.html). On that web page it mentions these files in the following:
“Write a file showing the mapping of each
node in the icosahedron to the closest
three nodes in the original mesh.
The file is named by the prefix of the output
spec file and suffixed by MI.1D”

I am aware that this is only one type of metric for alignment and other procedures, like using the Enigma QC, could be an alternative.

Could you please provide clarification about what these MI.1D files represent and whether they would be appropriate for evaluating alignment?

Thanks,
Cameron

Hi, Cameron-

Hmm, I see. I will have to dig into the *.MI.1D file a bit more, but my guess is that “MI” stands for “MapIcosahedron” rather than “mutual information”. It looks like it is more of a history of where it came from, listing nodes, rather than being a cost-like value. The full description of that option in the program help describes that a bit more:


-write_nodemap: (default) Write a file showing the mapping of each
                   node in the icosahedron to the closest
                   three nodes in the original mesh.
                   The file is named by the prefix of the output
                   spec file and suffixed by MI.1D
  NOTE I: This option is useful for understanding what contributed
        to a node's position in the standard meshes (STD_M).
        Say a triangle on the  STD_M version of the white matter
        surface (STD_WM) looks fishy, such as being large and
        obtuse compared to other triangles in STD_M. Right
        click on that triangle and get one of its nodes (Ns)
        search for Ns in column 0 of the MI.1D file. The three
        integers (N0, N1, N2) on the same row as Ns will point
        to the three nodes on the original meshes (sphere.reg)
        to which Ns (from the icosahedron) was mapped. Go to N1
        (or N0 or N2) on the original sphere.reg and examine the
        mesh there, which is best seen in mesh view mode ('p' button).
        It will most likely be the case that the sphere.reg mesh
        there would be highly distorted (quite compressed).

Looking at the top of a *MI.1D file in the AFNI Bootcamp data, the columns are described in comment as follows:


# Col. 0: Std-mesh icosahedron's node index.
# Col. 1..3: 1st..3rd closest nodes from original mesh (rh.sphere.reg.gii)
# Col. 4..6: 4th..6th interpolation weight for each of the 3 closest nodes.

So, again, I think this is more of a “mapping” than a “cost-like quantity”. (The “interpolation weight” is likely the information of relative proximity.)

In terms of wanting to quantify the accuracy of alignment of surfaces, that sounds like a good thing to have. But note that a cost function is a quantity that evaluates the overall quality of matching of 2 volumes: it is compared with alternative alignments during the processing, and in the end the arrangement with minimal cost wins and that is the final image. So, the final cost value itself is the quantitative assessment of matching. If there were a better quantity to assess the alignment, then by definition that quantity should be used as the cost function in the alignment process. And then, by definition, any other quantity to assess the matching must be a poorer evaluation than the cost function itself. Basically, it is inherently hard to quantitatively assess the quality of cost function-based alignment results, because you end up trying to replace the cost function: if you do find a better quantity to assess, then that would become the cost function to use.

In general, it seems like visual verification is still the gold standard for judging alignment, in the end. There are programs/functionalities in AFNI and SUMA to help visualize surface results systematically-- those might be of use?

–pt

Thank you so much for the prompt clarification. This is very helpful and helped us avoid a big mess up! Thanks again!

Sure thing, and happy to discuss any further points related to this/your project.

–pt