Use contrast-agnostic soft segmentation to improve atrophy measure

[PaulBautin]

Context

👉 Paper
👉 Documentation

A potential amelioration to the current segmentation method is the use of a soft mask that accounts for partial volume effect during segmentation. This method might improve sensitivity of CSA detection and by consequence improve detection of small atrophies. Additionally csa-atrophy, once T1 and T2 models are ready, could switch to new sct_deepseg,

Todo

• develop method to output non-binary segmentation 👉 https://github.com/sct-pipeline/contrast-agnostic-softseg-spinalcord
• re-run analysis for this study
• Investigate the 'contrast agnostic' feature of the new model
• re-run statistical analyses from the paper

Questions

• How to compute CSA from the soft mask? --> use sct_process_segmentation
• Is there a common threshold that should be used?

[jcohenadad]

How to compute CSA from the soft mask?

same procedure-- if you look at the doc of sct_process_segmentation, it takes soft inputs

Is there a common threshold?

no threshold ("soft" means non-binary)

are CSA voxels area weighted in sum to reflect uncertainty?

the other way around: uncertainty is reflected by soft voxels. Then, sum (used to compute CSA) is weighted

[jcohenadad]

The recently trained contrast agnostic model would be a perfect candidate for this project.

[sandrinebedard]

Update on this issue:

• Running on joplin takes arround 5 days, so I tried on compute canada
• Generates more than 651G of data...

Compute Canada

• Graham only has down to python 3.9, the code was at python 3.7, it was updated : see branch https://github.com/sct-pipeline/csa-atrophy/tree/sb/update-for-ca-python3.9
• Installing SCT on Graham: Clarify steps for installing SCT onto ComputeCanada Graham (requirements.txt conflicts) spinalcordtoolbox/spinalcordtoolbox#4593
• To many QC reports were generated at the same time, files were locked, I removed all the QC reports
• Reading and writing time are very long on graham, running inference takes 30sec, while the entire deepseg function takes 2 min (on joplin, 6s inference, entire function 30s)
• 1 subject takes 3h on joplin vs 1 day on graham (32 cores, 128 GM RAM...)

Next steps

• Run statistics on T2w results
• Run the analysis on T1w images
• Rerun the analysis without binarization for T2w and T1w