Dilation of spinal cord mask to remove lesions outside the spinal cord

[plbenveniste]

In a previous conversation with @jcohenadad and @tomDag25 we mentioned that when removing lesions outside of the spinal cord, the spinal cord mask should be dilated by 2mm on the axial plane.

I am wondering how to implement this since, sct_maths -dilate only takes an integer as input. Am I missing something? Is there an efficient way of doing it?

Tagging @mguaypaq and @joshuacwnewton for their input.

[joshuacwnewton]

I am wondering how to implement this since, sct_maths -dilate only takes an integer as input. Am I missing something? Is there an efficient way of doing it?

Looking at the -help text:

  -dilate <int>         Dilate binary or greyscale image with specified size. If
                        shape={'square', 'cube'}: size corresponds to the length
                        of an edge (size=1 has no effect). If shape={'disk',
                        'ball'}: size corresponds to the radius, not including
                        the center element (size=0 has no effect).

Given that you are looking to dilate along the axial plane, I believe shape=disk will do the trick? (It corresponds to skimage.morphology.disk, which I assume will span the x-y radius but not the z radius.)

However, I do note that you want to dilate in mm rather than px. So, you may need to do a quick conversion using the resolution of your image, since I believe -dilate only accepts pixel values.

[plbenveniste]

Thanks! I didn't understand how the shapes could be used with sct_maths -dilate.
And yes, for the dilation in mm, the conversion is not a problem.

[plbenveniste]

Actually, the dilation is not that straightforward!
Since the dilation only takes one argument, it cannot work for anisotropic images.

Example:

• image resolution: 1mmx2mm
• I want to dilate by 2 mm.
  In this case, I want a factor 2 for the x coordinate and a factor 1 for the y coordinate.

Am I missing something?

Example of command: sct_maths -i sc_seg_t2_raw.nii.gz -dilate 1 -shape disk -o dilated_mask.nii.gz -dim 2

Example for dilation by 1 or 2 voxel:

[jcohenadad]

right, I think that's currently a limitation of our dilation API. We should probably offer an option to specify the physical dimension (eg mm), in case in-plane resolution is anisotropic.

Alternatively, we could allow to specify each dimension. Eg: 1x2x0.

[plbenveniste]

Okay that's what I figured as well.

For the time being, I found a dirty solution so that I can perform the dilations for the challenge.

Since my images are 2D sagittal images, I can do the following:

1. I perform dilation on the axial plane by X (using a disk shape) to dilate the mask in the R-L direction.
2. I perform dilation on the sagittal plane by Y (using a disk shape) to dilate the mask in the A-P direction.

After step 1:

After step 2:

Eventhough, this will dilate on the I-S axis, I don't mind it since the goal is to eliminate lesions, not to refine segmentations.

[jcohenadad]

smart move