Data iteration: How to use vertebral labels with nnunet ?

[NathanMolinier]

Description

Using our current vertebral labels with the nnunet represents a challenge. Indeed, several issues and technical problems have to be considered:

1. Vertebral labels represent ground truth masks with a single coordinate for each intervertebral disc visible in the 3D corresponding image.
2. Additional labels beyond discs may be present in the ground truth masks, including the pontomedullary groove (label 49) and pontomedullary junction (label 50).
3. Vertebral labels were created from a 2D perspective (sagittal view) therefore impacting the accuracy of the labels in the left-right direction. (manual correction/ SCT viewer issue)

To cope with these problems, several scenarios will be investigated in this issue

[NathanMolinier]

Scenario 1

Solutions investigated:

1. Original vertebral labels will be replaced by new labels where the single pixels corresponding to each vertebral discs will be replaced by bigger spheres centered on this specific pixel:
   
   On this GIF you can see the red dots corresponding to the single pixels and the blue sphere corresponding to the new masks. Spheres have a radius equal to 4mm.

2. Labels with a number greater than 25 were removed during the creation of the new labels.

3. For this first iteration the new created labels from sct-testing-large were used. Special efforts were made on most of the images to ensure that labels are centered on the discs, the width of the spinal cord was used to this end.