RandCropByPosNegLabel: Random location

[lukasfolle]

Is your feature request related to a problem? Please describe.
I noticed that RandCropByPosNegLabel always crops objects with the foreground pixels being in the center.

Describe the solution you'd like
To get more variance into this transformation, I think it would be nice to have a flag indicating that the foreground voxel has not to be sampled in a centered way, but randomly in the patch. Especially for segmentation task with very few foreground voxels this would make sense IMO.

I am not entirely sure how that would affect model experiements but I would like to find out in case this gets implemented.
BTW: I absolutely love the functionality that this transformation already has!

[Nic-Ma]

Hi @lukasfolle ,

Thanks for raising this up.
Actually, we have a duplicated ticket: #586
Could you please help submit a PR for this new transform if you are interested?

Thanks.

[wyli]

i think this could be achieved by providing a dilated version of the original foreground as the sampling candidate locations (label_key). it’s slightly different from #586

[lukasfolle]

@wyli Could you elaborate on "dilated version of the original foreground as the sampling candidate locations"?

I had the following in mind:

1. Get foreground indices like before
2. Create a region around the foreground indices such that every valid crop yields a least one foreground index
3. Randomly sample the region

Please let me know what you think of this proposal.

[wyli]

yes, to achieve the same without modifying the transform:

1. dilate the foreground by k voxels and put it into the data dict with "label_dilated" key
2. invoke RandCropByPosNegLabel (label_key="label_dilated", spatial_size=2*k, ...)

[lukasfolle]

I agree on 1., but why should the spatial_size be changed in any way? If we choose k appropriately (depending on foregorund extent), shouldn´t the spatial_size be kept the same?

[wyli]

sure, dilation k and spatial size are free parameters here, I just wanted to set them in this particular relation so that the sampling roughly achieves "not to be sampled in a centered way, but randomly in the patch."

[lukasfolle]

@wyli Do you have some code working already?