Train 3D nnU-Net on segmentation results of 2D nnU-Net

[plbenveniste]

• Pre-process and transform data to the nnU-Net format
• Train and test 3D nnU-Net
• Visual analysis of the results
• Comparison with the results of the 2D nnU-Net

Related to #36

[plbenveniste]

• Created a python script to convert predictions back to the zurich_mouse dataset format
• Uploaded data on the romane server
• Currently running 3d_fullres training on romane gpu cluster 3 (estimated end time: June 20th at 10 am)

[plbenveniste]

Training completed:

• Trained on all data using segmentations from 2D nnU-Net trained on extracted slice
• 1000 iterations
• 5 folds
• 3d fullres
• Pseudo-Dice: 0.9

Observations (Comparison with segmentations using 2D nnU-Net):

• n°1, 2, 3, 5, 7, 8, 9, 10, 12, 13, 15, 18, 20: Similar result: new segmentation is more stable and contains less "noise"
  Example (image n°5):
  
  It shows that the annotations are more homogenous from slice to slice: advantage of going from 2D to 3D.

• n°4, 6, 11, 14, 17, 19: bad results in both case
  Example (n°6)
  
  The model learned to segment outside of the spinal cord.

General observation:

• in segmentation at the top or bottom of the volume of slice, the model tends not to label the entirety of the white matter. (the following example comes from n°14)
  

Suggestion for further improvement:

• Look at correlations between the type of chunk and the quality of the segmentation (cf image from bad example: usually this type of volume)
• Retrain the 2D model with more annotations on slices at the top/bottom of the volume to improve the quality of the segmentation in these areas and more annotations on types of volume which fail
• Observe the results ans select good annotations
• Retrain 3D model on these annotations only
• Observe results and conclude