TotalSeg configs and supporting code

[surajpaib]

No description provided.

[ibro45]

potentially problematic for new models

[surajpaib]

I think this won't be missed as newer models we'll have to figure out input configurations again and edit this as well. I'll move it over to a dict or so then

[ibro45]

let me know when i can merge this, will add head abnormality classification then following the same structure

[surajpaib]

@ibro45 You can merge it already if you like.

But I will add a script file for TS with specifics on how to run it.

Some changes I made,

name and project will be passed through scripts CLI overrides instead of in the config. There are way too many options now adding to confusion potentially.

[surajpaib]

@ibro45

I've tested all the configs once and looked added several improvements to the pipeline for TS

1. Val is on 256 cubic, this covers a lot more across the z-axis, see image

2. Metrics are generalized dice from MONAI fork (Fix generalized dice computation Project-MONAI/MONAI#7970). The main reason for adding this is that if pred is zero and gt is zero, the dice is set to 1. Torchmetrics sets this to zero and is difficult to override. Metrics are averaged across all labels and metrics are logged also per label with their semantic naming.

3. Added wandb image logger as seen in the image, its a simple callback on the first image/batch in the validation. This is mostly for sanity checks.

4. Data loading modified for TS so that different splits and label groups can be loaded based on dataframe filtering.

5. Predict workflow added (it runs but need to test for accuracy of results)

[surajpaib]

Going to run the comparison against Merlin now (end-to-end) to see if the workflow is good. I'll also move on to MSD and hope to finish that asap

[surajpaib]

@ibro45 Time to merge?

[surajpaib]

#10