Input Images that are fed into the model during prediction aren't normalized?

[jchen42703]

The model sees z-score standardized images during training, but not during evaluation. I think this should be fixed.

[jchen42703]

Wait. I think I made an incorrect assumption...The models aren't even trained with z-score standardized data?? This should be added to the preprocessing pipeline.

[jchen42703]

Let's retrain:

• kitsnnunet2dcrop_preprocessed_sparse_cnnfold1epoch80.h5
• kitsnnunet2dcrop_preprocessed_cnnfold1epoch15.h5 (sigmoid) while we're at it

[jchen42703]

Estimated Dataset-wide means:

• averaging the stats per_image:
  • 201 samples
  • Average mean: -29.2021338643; average std: 82.1537525668
• averaging the stats slice-wise and per_image combined:
  • 27673 samples
  • Average mean: -29.0928161528; average std: 82.1416007618

[jchen42703]

Fixed