[Neuroimaging] Analyzing the topology of ROIs and flood-filling in python (skimage?)
jomaroceguedag at gmail.com
Thu Sep 3 03:16:14 CEST 2015
I have been working on this issue for some days now (this is very
interesting Ariel!, thanks for sharing your findings). Satra is totally
right that **in theory** the transformations should preserve the topology.
Unfortunately, the transformations are only **approximately**
diffeomorphic. I am totally sure that this issue should be there in the
original version of ants too (dipy's implementation is the same algorithm),
although maybe the new version (antsRegistration) may have some
improvements that I'm not aware of.
Having said that, you can make the transforms closer to diffeomorphic by
reducing the `step_length` parameter (in millimeters) from
`SymmetricDiffeomorphicRegistration`, which by default is 0.25 mm. You may
try something about 0.15 mm. The objective is to avoid making very
"aggressive" iterations, so another way to achieve this is by increasing
the smoothing parameter from the CCMetric, the parameter is `sigma_diff`,
which by default is 2.0, you may try something bout 3.0 (I would first try
reducing the step size, though).
I would like to try some other ideas, by any chance can you share the data
Thank you very much!
On Wed, Sep 2, 2015 at 7:28 PM, Satrajit Ghosh <satra at mit.edu> wrote:
> hi ariel,
> can you do nearest neighbor interpolation in `mapping.inverse_transform`?
> if your original ROI doesn't have holes and you are doing a diffeomorphic
> mapping, your target shouldn't have holes either. for a comparison you
> could run antsRegister and antsApplyTransforms, with nearest neighbor
> On Wed, Sep 2, 2015 at 8:13 PM, Ariel Rokem <arokem at gmail.com> wrote:
>> Hi everyone,
>> Jason and I are working on a port of his AFQ system (
>> https://github.com/jyeatman/afq) into dipy. We've started sketching out
>> some notebooks on how that might work here:
>> The main thrust of this is in this one:
>> The first step in this process is to take a standard ROI of some part of
>> the brain (say, corpus callosum, which is where we are starting) and warp
>> it into the subject's individual brain through a non-linear registration
>> between the individual brain and the template brain on which the ROI was
>> defined (in this case MNI152). Registration works phenomenally (see cell
>> 17), but because this is a non-linear registration, we find ourselves with
>> some holes in the ROI after the transformation (see cell 27 for a
>> sum-intensity projects). We are trying to use
>> scipy.ndimage.binary_fill_holes to, well, fill these holes, but that
>> doesn't seem to be working for us (cell 35 still has that hole...).
>> Any ideas about what might be going wrong? Are we using fill_holes
>> incorrectly? Any other tricks to do flood-filling in python? Should we be
>> using skimage?
>> Neuroimaging mailing list
>> Neuroimaging at python.org
> Neuroimaging mailing list
> Neuroimaging at python.org
"Cada quien es dueño de lo que calla y esclavo de lo que dice"
"We all are owners of what we keep silent and slaves of what we say"
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Neuroimaging