Dear organisers,
I have a question regarding the expected output of the BraSyn challenge.
The goal is to generate the missing modality (missing MRI).
Is it expected the generated modality to be denormalised, i.e., should it be in the "real intensity values"?
Thanks!
Created by André Filipe Sousa Ferreira ShadowTwin41 Follow-up: The input folder looks like this:
/input-folder-example/
└── BraTS-GLI-01667-000/
└── BraTS-GLI-01667-000-t1n.nii.gz
BraTS-GLI-01667-000-t2f.nii.gz
BraTS-GLI-01667-000-t2w.nii.gz
└── BraTS-MEN-01346-000/
└── BraTS-MEN-01346-000-t1c.nii.gz
BraTS-MEN-01346-000-t2f.nii.gz
BraTS-MEN-01346-000-t2w.nii.gz
└── BraTS-MET-00777-000/
└── BraTS-MET-00777-000-t1c.nii.gz
BraTS-MET-00777-000-t1n.nii.gz
BraTS-MET-00777-000-t2f.nii.gz
The structure of the output folder is:
/output-folder-example/
└── BraTS-GLI-01667-000-t1c.nii.gz
BraTS-MEN-01346-000-t1n.nii.gz
BraTS-MET-00777-000-t2w.nii.gz
hi André @ShadowTwin41,
Sorry for my late response.
This year, we have switched to Docker containers.
Please predict the missing modality to a folder (without sub-folders).
Please see the updated tutorial: https://github.com/hongweilibran/BraSyn_tutorial
Best,
Bran
Dear @branhongweili,
What is the expected output structure? Should it predict the missing modality to a folder, or should it be inside of each patient subfolders?
Best,
André hi @ShadowTwin41,
To avoid potential issues, it's better to normalize the intensity to be between [0, 1] or [0, 255] (if your current output x is between [-1, 1], please do x = (x+1)/2).
Best,
Bran (Hongwei) Dear @branhongweili,
Thank for your answer.
Just to make sure, does it mean that the output can be normalized between -1 and 1?
@ShadowTwin41 Yes, I will update the tutorial here: [https://github.com/hongweilibran/BraSyn_tutorial](url) for local testing this Friday and the weekend. Please stay tuned.
Hi, for the SSIM metric, we re-normalize your generated image (and the ground-truth target image) to be [0, 1] after intensity clipping.
For the segmentation metric, the segmentation algorithm normalizes your data with z-score normalization (with a mean of zero and a standard deviation of 1). So it is not expected that the generated modality to be denormalised. Also, I would like to test my results using the DSC without using docker. Would it be possible?