Complete label sampler

[themantalope]

🚀 Feature

Sampler which is guaranteed to include all voxels with a certain label value.

Motivation

I am doing tumor localization/segmentation, and I need to sample patches of the images due to memory and spatial resolution constraints. I need to make sure that the tumors are within the sampled patches with a certain number of pixels/voxels between the edge of the tumor and the edge of the patch.

Pitch

Add a class to perform complete label sampling.

Alternatives

I've already rolled my own code to do this, but it might be useful for other folks who are working on similar things. Would be happy to contribute or figure out a way to add a class that is in line with the torchio framework.

[fepegar]

Hi, @themantalope. Thanks for the pitch. I'm not sure what you mean by "complete label sampling". Could you please elaborate?

[themantalope]

@fepegar

Apologies for not getting back for so long.

In my case, I am building pipelines for tumor detection/segmentation models.

The tumors can be varying sizes in my dataset, the one I'm working with has tumors ranging from ~1cc to 2000 cc in volume.

I wanted a transform that would allow me to randomly sample the image with a certain shape (let's say 128 x 128 x 128) but also ensure that the entire tumor was within the sampled image. The method I wrote allows you do do that, also specify what minimum padding you want from the edge of the volume and throws an error if the mask is too big for the volume requested (in my case that would mean at least one dimension of the tumor mask would be bigger than 128 pixels).