Add transform to ignore tumors smaller than x mm^3

[Nic-Ma]

Is your feature request related to a problem? Please describe.
Feature request from Ahmed: I just faced another interesting use case. Similar to keep largest connected component, Sometimes we want to ignore tumors that are smaller than x mm^3.
If you can implement this as a post transform it would be great.

[ebrahimebrahim]

Since I've been playing with ITK for these kinds of post processing tasks lately, I wanted to share that it's potentially an option here. It could allow for some flexibility/genericness with respect to the segment shape properties the user chooses to base their filter on: physical size, perimeter, centroid location, etc.

examples, assuming seg is a binary label map of shape (1,H,W):

import itk

# Convert tensor to ITK image
seg_itk = itk.image_from_array((seg.numpy()).astype(np.uint8))

# Get connected components
seg_connected_components = itk.ConnectedComponentImageFilter(seg_itk)

# Get a shape statistics object for each connected component
label_map = itk.LabelImageToShapeLabelMapFilter(seg_connected_components.astype(itk.UC))
num_label_objects = label_map.GetNumberOfLabelObjects()

# e.g. get the size of each connected component
[label_map.GetNthLabelObject(n).GetNumberOfPixels() for n in range(num_label_objects)]

# or various other properties
[label_map.GetNthLabelObject(n).GetPerimeter() for n in range(num_label_objects)]
[label_map.GetNthLabelObject(n).GetCentroid() for n in range(num_label_objects)]

[wyli]

MONAI/monai/transforms/utils.py

Lines 1002 to 1004 in c68e922

	def remove_small_objects(
	img: NdarrayTensor, min_size: int = 64, connectivity: int = 1, independent_channels: bool = True
	) -> NdarrayTensor:

the remove_small_objects has a min_size parameter which could be extended to take into account img.pixdim to compute mm^3 if the first argument img is a metatensor.

[rijobro]

Should we compute the volume from the number of voxels, or should we require the user to calculate it themselves? Either way it's a trivial calculation, but the large majority of our codebase is in image space (aside from Spacing-esque transforms). I would lean more towards staying in image space and requiring the user to calculate it themselves.

[vikashg]

closing because of inactivity and alternate solutions exist.