Better performance of deepseg seg_mice_gm-wm_dwi model if A-P orientation is wrong

[jcohenadad]

Description

While working on a user's forum issue https://forum.spinalcordmri.org/t/registration-to-t1-ex-vivo-template-error/644/6, I noticed that sct_deepseg's model seg_mice_gm-wm_dwi performs better if the image is (wrongly) flipped in the A-P direction.

See below with the correct orientation, ie: we clearly see the dorsal horns pointing towards "A" whereas it should be "P":

wrong-orientation.mov

However, when fixing the wrong orientation of the image using the following function:

sct_image -i dwi_1500.nii.gz -setorient-data PSL -o dwi_1500_orient.nii.gz

Even though the image is now correctly oriented, the segmentation performance is largely decreased:

correct-orientation.mov

Which makes me wonder if the model may have been trained on flipped data? (ie: A-P were mislabeled). @charleygros

[charleygros]

Which makes me wonder if the model may have been trained on flipped data? (ie: A-P were mislabeled). @charleygros

That's.... possible!!
What I'm gonna do: (1) flip the dataset, (2) retrain a model on this flip data, (3) compare the results, (4) make a call. Keep you posted.

[jcohenadad]

i was about to suggest you wait for the data to be RPI (instead of LPI) but since the template is RL symmetric that shouldn’t change anything

[charleygros]

i was about to suggest you wait for the data to be RPI (instead of LPI) but since the template is RL symmetric that shouldn’t change anything

Good point! But I think these data are not RL symmetric (while the template has been RL symmetrised).

[jcohenadad]

i was about to suggest you wait for the data to be RPI (instead of LPI) but since the template is RL symmetric that shouldn’t change anything

Good point! But I think these data are not RL symmetric (while the template has been RL symmetrised).

ah! indeed. But data augmentation probably includes RL flip?

[charleygros]

data augmentation probably includes RL flip?

Yess!!