search

CmrxRecon / CMRxRecon2024

Code for the Cardiac MRI Reconstruction Challenge 2024 (CMRxRecon2024)
https://cmrxrecon.github.io/2024/Home.html
16 stars 3 forks source link

Normalization of predictions #2

Open fzimmermann89 opened 4 months ago

fzimmermann89 commented 4 months ago

Please reconsider the normalization of the prediction you used last year (and is still done in the 2024 snippets repository)

If normalizing pred by pred.max, a single high Pixel in the predicted images completely destroys all metrics.

Either do not normalize and provide a single example of a gt validation result such that teams can ensure they use the correct normalization. Or maybe normalize by a more stable value, such as Std.

For us, option 1 would be prefered.

ziqiangxu commented 4 months ago

Hi, fzimmermann89, thank you for your reminder.

Please check these lines @Mobbyjj https://github.com/CmrxRecon/CMRxRecon2024-snippets/blob/c5cb30c775e2cfc62871fc71128852aac5f3178d/evaluation-2024/Evaluation.py#L51C1-L55C74

Mobbyjj commented 4 months ago

Hi Felix,

Thanks for bring this into our attention. We have revised the normalisation using the std: https://github.com/CmrxRecon/CMRxRecon2024-snippets/pull/4#issue-2287416949

Cheers!