Is dilation necessary at inference time?

[jcohenadad]

Dilation adds a line of code for users, so it would be nice if this step was not necessary. See e.g. this tutorial for using the new models: spinalcordtoolbox/spinalcordtoolbox#3637 (comment)

I've tried without and with dilation (ball=5) in one subject (P007), and results are slightly different, but a more systematic comparison should be done in order to advise on the best course of action.

@uzaymacar do you have any suggestions/insights? I'm thinking that maybe in the future we could also train without dilation (I don't think it is really necessary because the SC will always be included in the cropped image, even without dilation).

[joshuacwnewton]

Dilation adds a line of code for users, so it would be nice if this step was not necessary. See e.g. this tutorial for using the new models: spinalcordtoolbox/spinalcordtoolbox#3637 (comment)

Perhaps I'm missing something, but I don't see a dilation step in the linked comment. Could you clarify what you mean by "line of code for our users"?

[jcohenadad]

Sorry for the lack of clarity. Indeed, the link to spinalcordtoolbox#3637 did not include the dilation step, which is precisely the point of this issue.

By default, dilation is performed during preprocessing, before training the model:

model_seg_ms_mp2rage/preprocessing/preprocess_data.sh

Line 123 in 0dccf78

sct_maths -i ${file_seg}.nii.gz -dilate 5 -shape ball -o ${file_seg}_dilate.nii.gz

And my suggestion would be to get rid of this dilation step.

[uzaymacar]

I saw dilation as an non-vital, extra preprocessing step to make sure we capture the ROI during inference and model development. For example, I would expect it to be helpful in cases where lesions are near the boundaries of the SC mask.

That being said, the following next steps can help us decide whether or not we can remove the dilation step safely:

• Inference-time: Compare lesion segmentation results with and without dilation on all test subjects (extend @jcohenadad's work)
• Training-time: Train SC segmentation without dilation, then train lesion segmentation on top of the new SC-cropped images, and compare lesion segmentation results to before