Next steps

[jcohenadad]

in order:

• Merge Fixed missing top disc label with using -discfile  spinalcordtoolbox/spinalcordtoolbox#2990
• Currently Add new DL-based method (Countception) for C2-C3 disc detection and template matching  spinalcordtoolbox/spinalcordtoolbox#2679 is failing -- fix it
• Update README to be comprehensive. Ex: "sct_run_batch -c parameters/<CONFIG_FILE>" --> unclear what script to run (in what order, explain what it does, blablab)
• Run pipeline, evaluate results, if good create release 0.1 and upload results incl. exact version of dataset used, SCT used, etc.

[joshuacwnewton]

• Update README to be comprehensive. Ex: "sct_run_batch -c parameters/<CONFIG_FILE>" --> unclear what script to run (in what order, explain what it does, blablab)

This is done, I think. :)

• Run pipeline, evaluate results, if good create release 0.1 and upload results incl. exact version of dataset used, SCT used, etc.

I've been tracking TODOs for this in my personal TODO board, but just to give an update on what I've been working on + planning:

• Find out which metrics were used to evaluate previous deep-learning approaches. (Read Reza/Lucas' Countception, Hourglass papers)
• Brainstorm list of additional metrics to use, if needed
• Tweak Python script for computing metrics (existing evaluation script was broken)
• Fine-tune "smaller" version of sct-testing-large dataset (remove/add suitable subjects) for quick iteration when evaluating retrained models (existing subjects were not suitable for testing)
• Retrain Hourglass using same data augmentation as Countception, reevaluate
• Download full sct-testing-large suitable subjects, then run pipeline to compare models in-depth
• Present findings in future ivadomed-internal meeting