search

sct-pipeline / spine-park

Pipeline for multicontrast analysis in PD patients
MIT License
0 stars 0 forks source link

Vertebral labeling is unreliable on this dataset using `sct_label_vertebrae` #9

Closed jcohenadad closed 7 months ago

jcohenadad commented 9 months ago

Version: https://github.com/sct-pipeline/spine-park/pull/8/commits/9d780cc67cd15519db437dfc48c4878cdc55030b

We should explore alternative approaches. @NathanMolinier do you have a method that we can already use for this project?

Kaonashi22 commented 9 months ago

Version: 9d780cc

Is this vertebral labeling accurate? The masks don't correspond to the vertebral body

vertebral_labels

jcohenadad commented 9 months ago

Is this vertebral labeling accurate? The masks don't correspond to the vertebral body

no it is not, hence this issue 😉

Kaonashi22 commented 9 months ago

True! I guess the registration to the template is not accurate if the labels are wrong...

Kaonashi22 commented 9 months ago

This might be a different problem: the vertebral labeling and segmentation failed for this subject. There is a slight lateral curvature at the thoracic level (on coronal views); I don't know if it may impact the process

1 2

jcohenadad commented 9 months ago

This might be a different problem

Same problem. We are working on an alternative solution for vertebral labeling. Stay tuned 😊

NathanMolinier commented 9 months ago

Is the data available somewhere ?

Kaonashi22 commented 9 months ago

Thanks, I can give you some sample images. How can I send them?

From: Nathan Molinier @.> Sent: February 29, 2024 15:19 To: sct-pipeline/spine-park @.> Cc: Lydia Chougar, Dr @.>; Comment @.> Subject: Re: [sct-pipeline/spine-park] Vertebral labeling is unreliable on this dataset using sct_label_vertebrae (Issue #9)

Is the data available somewhere ?

— Reply to this email directly, view it on GitHubhttps://github.com/sct-pipeline/spine-park/issues/9#issuecomment-1971889331, or unsubscribehttps://github.com/notifications/unsubscribe-auth/BFCFJYVKDGP2CK46WTZ5ZKLYV6GLNAVCNFSM6AAAAABD5ITGAKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSNZRHA4DSMZTGE. You are receiving this because you commented.Message ID: @.***>

NathanMolinier commented 9 months ago

Maybe using dropbox or google drive ? Also, is it only a T2w contrast for all the images ? And the FOV is only cervical ?

Kaonashi22 commented 9 months ago

Can I have your email? I have T2, MT, M2RAGE and diffusion images for ~ 50 subjects The FOV includes the upper thoracic level until T5-T6

From: Nathan Molinier @.> Sent: February 29, 2024 15:48 To: sct-pipeline/spine-park @.> Cc: Lydia Chougar, Dr @.>; Comment @.> Subject: Re: [sct-pipeline/spine-park] Vertebral labeling is unreliable on this dataset using sct_label_vertebrae (Issue #9)

Maybe using dropbox or google drive ? Also, is it only a T2w contrast for all the images ? And the FOV is only cervical ?

— Reply to this email directly, view it on GitHubhttps://github.com/sct-pipeline/spine-park/issues/9#issuecomment-1971933250, or unsubscribehttps://github.com/notifications/unsubscribe-auth/BFCFJYV6A5OXYP7ZKE54PODYV6JZDAVCNFSM6AAAAABD5ITGAKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSNZRHEZTGMRVGA. You are receiving this because you commented.Message ID: @.***>

jcohenadad commented 9 months ago

Email introduction sent (I suggest to send the data privately)

NathanMolinier commented 9 months ago

Update

Hey ! Just to give a few feedbacks:

sub-BB277 sub-DC293 sub-DEV148Sujet01 sub-MC224
Kapture 2024-03-01 at 14 28 14 Kapture 2024-03-01 at 14 31 31 Kapture 2024-03-01 at 14 34 29 Kapture 2024-03-01 at 14 40 28

I also wanted to test with other contrasts, but they managed to catch the contrast in the filename ahah.

They are also using the BIDS standard

Screenshot 2024-03-01 at 14 57 53

I could still technically rename the files to bypass this check but I don't really expect much since their models were only trained with T2w contrasts.

Conclusion

It looks like we found a interesting model for vertebral labeling on T2w images ! Therefore, I will add this new method to the benchmark to evaluate its performances with different FOVs ! Also, for this project I could add an option to only extract the discs labels like we are doing with sct_label_vertebrae.

jcohenadad commented 9 months ago

It looks like we found a interesting model for vertebral labeling on T2w images ! Therefore, I will add this new method to the benchmark to evaluate its performances with different FOVs ! Also, for this project I could add an option to only extract the discs labels like we are doing with sct_label_vertebrae

Amazing! yes, the next step is indeed to find a way to extract the disc labels. Maybe you can open an issue on SCT website so we can brainstorm on how to do this best? Eg: what if we use the current 'project label on centerline' feature from sct_label_utils?

Kaonashi22 commented 9 months ago

Good news, thanks!

From: Julien Cohen-Adad @.> Sent: March 1, 2024 15:28 To: sct-pipeline/spine-park @.> Cc: Lydia Chougar, Dr @.>; Comment @.> Subject: Re: [sct-pipeline/spine-park] Vertebral labeling is unreliable on this dataset using sct_label_vertebrae (Issue #9)

It looks like we found a interesting model for vertebral labeling on T2w images ! Therefore, I will add this new method to the benchmark to evaluate its performances with different FOVs ! Also, for this project I could add an option to only extract the discs labels like we are doing with sct_label_vertebrae

Amazing! yes, the next step is indeed to find a way to extract the disc labels. Maybe you can open an issue on SCT website so we can brainstorm on how to do this best? Eg: what if we use the current 'project label on centerline' feature from sct_label_utils?

— Reply to this email directly, view it on GitHubhttps://github.com/sct-pipeline/spine-park/issues/9#issuecomment-1973868612, or unsubscribehttps://github.com/notifications/unsubscribe-auth/BFCFJYXP5XQ2RMA64B7PPSDYWDQF3AVCNFSM6AAAAABD5ITGAKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSNZTHA3DQNRRGI. You are receiving this because you commented.Message ID: @.***>

lydiachougar commented 8 months ago

Hi! Any updates regarding the vertebral labeling?

NathanMolinier commented 8 months ago
Hi ! I am still experimenting on this but I will provide you some updates and inference code next week ! Labeling based on discs segmentations Output
test1 test2
Kaonashi22 commented 8 months ago

Sounds good, thank you!

From: Nathan Molinier @.> Sent: March 8, 2024 09:23 To: sct-pipeline/spine-park @.> Cc: Lydia Chougar, Dr @.>; Comment @.> Subject: Re: [sct-pipeline/spine-park] Vertebral labeling is unreliable on this dataset using sct_label_vertebrae (Issue #9)

Hi ! I am still experimenting on this but I will provide you some updates and inference code next week !

Labeling based on discs segmentations Output test1.png (view on web)https://github.com/sct-pipeline/spine-park/assets/68945192/f745d65d-9c12-4766-b05a-2e647aa2f92b test2.png (view on web)https://github.com/sct-pipeline/spine-park/assets/68945192/64c80746-5257-4739-9f19-30eccc95fb7b

— Reply to this email directly, view it on GitHubhttps://github.com/sct-pipeline/spine-park/issues/9#issuecomment-1985783348, or unsubscribehttps://github.com/notifications/unsubscribe-auth/BFCFJYWXUT5EKVPFH7REVHLYXHCVBAVCNFSM6AAAAABD5ITGAKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSOBVG44DGMZUHA. You are receiving this because you commented.Message ID: @.***>

NathanMolinier commented 8 months ago

Sorry for the slight delay, but I'm actively working on a solution to make this vertebral labeling work on every contrast. For now, I can only provide you the script that I wrote, based on SPINEPS, to extract the discs of T2w images. (The output file with discs labels has the suffix _label-discs_dlabel in the output folder)

Regarding the environment I just followed the instructions from SPINEPS. Also, the inference only works using the GPU.

If you encounter any issues with my script please let me know, I will try to fix it as soon as possible.

Note: If the top of the spine is not visible in the image, the predicted labels won't be correct unfortunately. So be careful

Kaonashi22 commented 8 months ago

@NathanMolinier, thanks a lot! @jcohenadad, should I run Nathan's script separately or add it to the main preprocessing script?

NathanMolinier commented 8 months ago

I personally think that you should run my script separately to avoid having SPINEPS as a dependency ! A more stable version will be implemented into the spinalcord toolbox.

jcohenadad commented 8 months ago

OK, to makes things easier and reproducible, I think that we should update the processing script to call SPINEPS instead of sct_label_vertebrae.

To do this smoothly, the first step is for you to install SPINEPS and make sure it is callable from the Terminal. Once this is done, I will update the script to call SPINEPS.

Kaonashi22 commented 8 months ago

Sounds good, I'll let you know how it works

From: Julien Cohen-Adad @.> Sent: March 20, 2024 11:13 To: sct-pipeline/spine-park @.> Cc: Lydia Chougar, Dr @.>; Comment @.> Subject: Re: [sct-pipeline/spine-park] Vertebral labeling is unreliable on this dataset using sct_label_vertebrae (Issue #9)

OK, to makes things easier and reproducible, I think that we should update the processing script to call SPINEPS instead of sct_label_vertebrae.

To do this smoothly, the first step is for you to install SPINEPS and make sure it is callable from the Terminal. Once this is done, I will update the script to call SPINEPS.

— Reply to this email directly, view it on GitHubhttps://github.com/sct-pipeline/spine-park/issues/9#issuecomment-2009811110, or unsubscribehttps://github.com/notifications/unsubscribe-auth/BFCFJYSBH25JGNWRL3CVKLDYZGRP7AVCNFSM6AAAAABD5ITGAKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDAMBZHAYTCMJRGA. You are receiving this because you commented.Message ID: @.***>

Kaonashi22 commented 7 months ago

I finally succeeded to install SPINEPS. Hopefully it will be working well!

jcohenadad commented 7 months ago

Vertebral labeling will be done manually. See #26 #24 #21