Results on nnU-Net trained over 97 subjects mixing healthy and mild compression subjects

[abelsalm]

Here is a review on the last training I did on ann nnU-Net for spinal canal segmentation.
I added more data using the previous training, 10th issue, trained the model on 97 subjetcs using the results of the inference, mixing data from spinge-generic, dcm-oklahoma, dcm-brno, sci-paris, and dcm zurich.
On this training, I had at least 15 subjects from each dataset.

I trained the model for 500 epochs, with on a fold I chose :
splits_final.json
So here there were only 7 validations cases, I wanted the model to learn as much as possible just having a few heterogenous validation cases to see the progress.

Here is the training progress:

I didn't add testing cases since here computing a dice score on those segmentation doesn't really give information one the accuracy of the model for two reasons:

• Dice score is really high already so it is hard to really measure improvements (we're near 0.98 here)
• Most of the parts that do not correspond between the labels and the generated labels are on the most superior part of the canal, where the model sometimes hesitates where to stop. So it doesn't really give information, seeing it on an axial opoint of vue, on the correspondance between groundtruth segmentation slices and results by the model.

May be I'll try to generate a QC on the inference I runned for this @jcohenadad @sandrinebedard @valosekj

Do I have to also make a release for this model ?
Also I'll try to make a pull request on a new branch

[abelsalm]

Here is a QC I generated, I'm really happy with the results.
You'll may be notice that some segmentations of concavities are a bit missed, but it only appears on the QC: trying to verify it with ITK Snap I realised that most of those are in fact quite accurate, I don't know why it gives this impression on the QC report.

I can't upload the zipped QC since it's too big, would you want screenshots or can I send it with something else ?

[jcohenadad]

Here is a QC I generated, I'm really happy with the results.

where is it?

[abelsalm]

Here is a QC I generated, I'm really happy with the results.

where is it?
I didn't manage to upload it... I'll make screenshots today, or I can send it to you differently may be ?

[jcohenadad]

I didn't manage to upload it... I'll make screenshots today, or I can send it to you differently may be ?

zip it and upload-- if too big use another cloud storage server than github

[abelsalm]

Here is the QC !
https://drive.google.com/file/d/1-eIgL33GkPWcPPvxg97nLoRP6QOPuVy2/view?usp=drive_link

[sandrinebedard]

Looks good!

A few problematic:

• sub-003_ses-01_T2w_000_0000.nii.gz
• sub-3140B_ses-3140B_T2w_000_0000.nii.gz
• sub-3641B_ses-3641B_T2w_000_0000.nii.gz
• sub-864639_acq-axial_T2w_000_0000.nii.gz (oversegmented I think)
  A few of the last slices have holes like for sub-3230B_ses-3230B_T2w_000_0000.nii.gz
  

Since sometimes teh canal is cropped in th QC, would it be worth it modifying the QC (seg is used for center cropping) but then the cropping size is hard codded (I think)

Maybe this image is not worth keeping in your testing set?
sub-CSM081_T2w_000_0000.nii.gz, data quality is really bad...