Ensure that labeling is reliable when the FOV is only thoracic

[NathanMolinier]

Description

The aim of this issue is to validate the relevance of the labeling when the following classes are not visible in the image:

• The sacrum (92)
• The disc L5-S1 (202)
• The vertebrae C1 (41)
• The vertebrae C2 (40)
• The disc C2-C3 (224)

Method

To check this information, 30 images were randomly selected from 3 different datasets unseen during training.

1. The inference was run on these images a first time to identify different classes:

   • The sacrum (92)
   • The disc L5-S1 (202)
   • The vertebrae C1 (41)
   • The vertebrae C2 (40)
   • The disc C2-C3 (224)

2. The same images were then cropped using the generated segmentations to exclude the top and bottom parts of the images where these previous classes were visible.

3. The inference was run another time but this time using the cropped versions of the images.

Results

commit 4ba0d78

Each white square presents the original segmentation (left) and the segmentation after cropping.

Discussion

1. In some images, we can observe that there is a shift between the discs and the vertebrae probably due to fact that the:

• the discs labeling relies on discs landmarks
• the vertebrae labeling rely on vertebrae landmarks
  Will be fixed in Refactor iterative labeling logic init from discs only #54

2. When the FOV is really small (axial aquisition) the model is not able to perform the labeling correctly, so no output is ultimately generated. Using a localizer to help the labeling as proposed in this PR is the solution to fix this issue.

[NathanMolinier]

Results

commit 253ce5a

These results were created based on this #54

[yw7]

Hi @NathanMolinier ,
Thank you for this amazing investigation!
Is it possible to include also the raw results (step2_raw) so we can have more details about the cause for the problem?

[NathanMolinier]

Yes I will do that for next iterations !