Cadotte et al 2015 prediction problems

[tzebre]

I have tried to predict the nerve segmentation of the article images but the results are really bad. C2-C3 are labeled some T1-T2 also but a lot of images have no segmentation on the middle levels.

A possible explanation, the resolution is .3mm model has only been trained on 0.6 and 0.8. On a 0.3mm image you can see individual rootlets:

I have tried to oversample the image to .7mm before the prediction. The same problem appears middle spinal levels are not labeled, but it looks like C2-C3 levels have a better labeling.

[jcohenadad]

I have tried to oversample the image to .7mm before the prediction.

You meant downsample (oversample is going from low res to high res)

[valosekj]

Predictions seem to be better with M5 fold_all model:

--> closing