Crop to bounding box of the mask

[mattwarkentin]

Hi @fepegar,

Is there a way to crop to the bounding box of the non-zero values in a mask?

Such as with data like:

import torchio as tio

sub = tio.Subject(
  image = tio.ScalarImage(...),
  mask = tio.LabelMap(...)
)

# Functionality like
tio.Crop('mask')(sub)

[romainVala]

Hello
No currently, only the CropOrPad transform can take a mask_name argument in addition to target_shape
The center of mass of the mask is then used to choose the center of the FOV (field of view) of the corped or paded volume
I thing it should do the job, you just have to guess the final target shape you wish

If you implement, it , I would go with CropOrPad, but the risk is then if mask are changing from subject to subject, you may end up with different crop between subject and thus a different shape, which is not a good idea

I hope it helps

[fepegar]

This seems to be related to #675

@mattwarkentin, can you tell us a bit more about what you're planning to do?

[mattwarkentin]

you just have to guess the final target shape you wish

Indeed this is what I am trying to avoid. By choosing a fixed size, you will end up with more or less (perhaps undesired) volume surrounding the ROI.

but the risk is then if mask are changing from subject to subject, you may end up with different crop between subject and thus a different shape

This is true, but I think when paired with the recently added tio.Resize transform it can work well.

can you tell us a bit more about what you're planning to do?

@fepegar I think there are use-cases where one might want to crop an ROI to "fill" the FOV. In other words, even if the ROI is a different size across subjects, the cropped volume will have the ROI centred and generally filling the cropped volume space. In the example of #675, one may wish to crop down to the bounding box of the lung mask, and then resize so that the output volumes for all subjects are equivalent in shape.

For example,

sub = tio.Subject(
    image = tio.ScalarImage(...),
    mask = tio.LabelMap(...)
)

trans = tio.Compose([
    tio.CropToMask('mask'),
    tio.Resize((100, 100, 100))
])

sub_t = trans(sub)

In this case, you will always get a consistent volume size that contains the entire ROI for every subject; and the ROI will generally fill the viewport. Hopefully that makes sense. Please let me know if you think this is not sensible.

[fepegar]

Hmm. Are you requesting support to specify the labels in the label map to compute the ROI?

[mattwarkentin]

I suppose. If we are cropping down to a bounding box, the bounding box could be for a single label in the label map or for a set of labels (similar to the select_fn in CropForeground).

[fepegar]

Alright! How's this?

import torchio as tio
cp = tio.CropOrPad(mask_name='brain')
colin = tio.datasets.Colin27()
colin_tight_brain = cp(colin)
colin_tight_brain.plot()

import torchio as tio
cp = tio.CropOrPad(mask_name='cls', labels=[3])
colin = tio.datasets.Colin27(2008)
colin_tight_wm = cp(colin)
colin_tight_wm.plot()

[mattwarkentin]

Wow, that was fast. LGTM!

[fepegar]

Kudos to @GReguig, @efirdc and @Svdvoort, who contributed the modular features in CropOrPad, the labels transforms and Mask that makes these changes easy.

I will open a PR in a second.