Visualization: Compare best,median,worse cases for all benchmarked methods

[msrepo]

• bring in code from MeshVisualizer that uses fury and xvfb
  try thin instead of vedo because of aesthetics
• use filter_runs_from_wandb to find run-id for each anatomy and method under consideration
• find evaluation/metrics.csv and find the best, median and worse case sample-id

[msrepo]

pandas has a weird API

print(df.nlargest(1, ["DSC"], "first")["subject-id"].values[0])
print(df.nsmallest(1, ["DSC"], "first")["subject-id"].values[0])
print(df[df.DSC == df.median(numeric_only=True)["DSC"]]["subject-id"].values[0])

[msrepo]

Median Results

[msrepo]

pick random samples instead of median, worse, best.
currently, the results are stored as follows:

keep the same configuration, but we need a way to identify random samples for tiling later.

def save_montage(ANATOMY, subject_type):

will also need to be modified. either track the subject-id or find a way to consistently track the random samples by consistent naming convention.

[msrepo]

Useful comparative visualization

[msrepo]

Median Results

The prediction resolution resampling and metadata issue ,see #1, has risen its ugly head here. I must have changed the parameters when evaluating attentionunet results for vertebra.

[msrepo]

Median Results

The prediction resolution resampling and metadata issue ,see #1, has risen its ugly head here. I must have changed the parameters when evaluating attentionunet results for vertebra.

Without resampling, metadata copied from groundtruth

without resampling, without metadata copied from groundtruth

both of these are wrong.

1. either set resample=True when you copy metada from groundtruth
   
   The preview happened correctly even if the data and metadata did not match in this case because we reslice when previewing.
   

2. or set resample=False and set the correct metadata manually.

[msrepo]

fixed visualization issues: see #28
before and after