Hi,
I have some questions regarding the task and the test submission:
- Will we receive T1,T2,Flair from the same subjects and can we assume all three exist at test time? That is, could we predict high field Flair by using all three low field modalities?
- Will we know at test time which modality in the input is which? For example, could we change some settings when predicting Flair?
- are all subject healthy volunteers or does the test set contain patients?
- can you provide the exact metrics you will use? Will you do any normalization or masking? What settings (window, window size etc) will you use for SSIM?
Thanks for organizing,
Cheers
Created by anon Anon trotteligerotter Hi @trotteligerotter, thanks for your reply. The validation and test dataset structures will be exactly like the training dataset. There will be no cases of missing contrasts. About the "a single outlier pixel will change the normalisation and thus the MSE of the whole image", I will discuss this with the team and get back to you. Thank you. Thank you for these pointers.
I still don't understand fully how the final scoring is done: we must submit a docker container that somehow be called to generate our predictions from the low field input data. We must save as nifti, then you provided evalöuation script will calculate the metrics.
The questions remains if we can use all modalities of a subject to jointly predict the high field results. Will at test time the container gut access to a folder in similiar structure as the training set, where for each subject we get all three modalities?
Or will there be some randomization, missing data, etc to force us to only use one of the modalities to predict the corrosponging high field image?
Will you reconsider the normaisation done in the evaluations?
Currently a single outlier pixel will change the normalisation and thus the MSE of the whole image. Can you instead use mean/std normalisation to have a more stable metric?
Cheers
Hi @trotteligerotter, thank you for your message. Please consider reading the challenge document at https://zenodo.org/records/15259777 and our GitHub repository at https://github.com/BioMedAnalysis/ULF-EnC-Challenge. If you still have any confusion or need further clarification, please feel free to message us here. Thank you for your participation.