change default pading for RandomAffine to minimum

[romainVala]

🚀 Feature
I notice that by default you choose the minimum as default_pad_value for elastic deformation
whereas the default method is otsu for randomAffine

Motivation
it seems to me more natural to take the min, as noise is expected in the image border (at least for brain, in 5 of the 6 border slices )
no ?
and it makes more sense to have the same default for both transform

Pitch
by the way we should may be add a default_pad_value as argument in the randomElastic too (but minimum is fine, so may be latter ...)

Alternatives

Additional context

[fepegar]

it seems to me more natural to take the min, as noise is expected in the image border (at least for brain, in 5 of the 6 border slices )

Sometimes the minimum is very different from the border, so it creates a gradient. I feel like using the Otsu threshold of the values at the border is the most robust option. But I can't think of an obvious counterexample at the moment...

and it makes more sense to have the same default for both transform

Yes! I added the padding stuff to RandomAffine as an experiment, but it should also be added to Resample and RandomElasticInformation. Actually, it would be nice to merge these 3 one day.

[romainVala]

adding the mean value (of the otsu mask) will also add gradient in the image borders
may be padding stuff is a proper solution (I did not know it was possible, does it allow to choose 'reflect' ? )

When talking of the same default I was thinking of changing RandomAffine default default_pad_value, which is now 'otsu' whereas it is 'minimum' in RadomElasticInformation (without the possibility to change it)

Not sure what you have in mind for merging those 3 transforms, but a first step would be to homogenize them, by adding the default_pad_value and the padding stuff if possible (with the same default)

[fepegar]

adding the mean value (of the otsu mask) will also add gradient in the image borders

I guess.

may be padding stuff is a proper solution (I did not know it was possible, does it allow to choose 'reflect' ? )

Hmm you can always use Pad with any mode you want, then apply the affine and then Crop.

When talking of the same default I was thinking of changing RandomAffine default default_pad_value, which is now 'otsu' whereas it is 'minimum' in RadomElasticInformation (without the possibility to change it)
Not sure what you have in mind for merging those 3 transforms, but a first step would be to homogenize them, by adding the default_pad_value and the padding stuff if possible (with the same default)

I agree, this needs to be changed. The idea of merging is avoiding loss of information due to multiple interpolations.

[romainVala]

Yes! I added the padding stuff to RandomAffine as an experiment

then I do not understand what you add ?

[fepegar]

Maybe "padding" (used by e.g. NiftyReg) is not the right nomenclature. I meant that I added logic to compute the "default pixel value" (used by ITK) when no information is available for that voxel during resampling. Does that make sense?

[romainVala]

sorry, but not sure to follow

when no information is available for that voxel during resampling.

this is already handel (in RandomAfinne)

        default_pad_value: As the image is rotated, some values near the
            borders will be undefined.
            If ``'minimum'``, the fill value will be the image minimum.
            If ``'mean'``, the fill value is the mean of the border values.
            If ``'otsu'``, the fill value is the mean of the values at the

do you mean you add other logic than minimum, mean otsu, or constant value ?, which logic then ?

[fepegar]

I don't know what the question is anymore haha.

When you apply e.g. a rotation to an image, you need to pad the result with some values. For example, in this image, the value is 0:

The minimum could have been chosen instead, or some other value.

That's all. I'm not proposing any new technique to compute the padding value. I agree the method should be the same for RandomAffine, RandomElasticDeformation and Resample.

[romainVala]

Yes! I added the padding stuff to RandomAffine as an experiment,

it is just this sentence I did not understand (since it is already there, in random Affine), but ok, no matter, I think we do agree (and it is not a very important point)

[fepegar]

I meant that back then I added that to RandomAffine and forgot to add it to the other transforms as well.

[fepegar]

Closing as this was fixed in #353. Next step should be adding padding options to Resample and RandomElasticDeformation.