SC lesion comparison between M0 and M12

[plbenveniste]

In this issue, we investigate the patient's differences between M0 and M12.

To do so, we first listed all the patients which had to be excluded, which are :

• patients with poor PSIR or STIR data quality
• patients with impossibility to label vertebral levels
• patients which are not present and either M0 or M12
• patients which have the same images at M0 and M12 (issue 39 for ref

The following file contains all the subjects to be excluded : exclude_M0_M12_comparison.yml

TO DO:

• Run inference on M12 subjects
• Generate dataframe for M0 subjects
• Generate dataframe for M12 subjects
• Compare dataframes

[plbenveniste]

Inference was run on M12 time points with both the 2d and 3d nnUNet (trained on STIR and multiplied by -1 PSIR) → inference was performed on 290 subjects.
Inference results are available at :

• duke.neuro.polymtl.ca/temp/plben/inference_results_555_2d_BIDS
• duke.neuro.polymtl.ca/temp/plben/inference_results_555_3d_BIDS

The dataframes were generated for M0 and for M12 3d and M12 2d. They include every subject in the CanProCo except from those in exclude_M0_M12_comparison.yml They are available at :

• duke.neuro.polymtl.ca/temp/plben/dataframe_M0
• duke.neuro.polymtl.ca/temp/plben/dataframe_M12_2d
• duke.neuro.polymtl.ca/temp/plben/dataframe_M12_3d

→ Currently working on dataframe comparison

[plbenveniste]

The analysis of M0 and M12 dataframes was performed.
→ figure to be used for abstracts ?

However, I noticed, that the total volume of lesion segmented is greater with the 2d nnUNet inference then with the 3d nnUNet inference.

[jcohenadad]

→ figure to be used for abstracts ?

for ismrm definitely (not possible for ACTRIMS AFAIK)