hongweilibran / brasyn

materials for the BraSyn challenge at MICCAI

---

This is a tutorial about BraSyn challenge's setup and specific implementations during the validation and test stages.

Data format of the test set

Each patient in the test set has the same naming format as the ones in the training and validation sets.
However, one modality will be randomly dropped during the test stage.

We provide a Python script named dropout_modality.py to show how it works on the validation set.

Please note that your container will only take on one folder as the input and be iterated on the whole test set. You will expect the format like this:

   BraTS-GLI-99999-000

    |  BraTS-GLI-99999-000-t1c.nii.gz

    |  BraTS-GLI-99999-000-t1n.nii.gz

    |  BraTS-GLI-99999-000-t2f.nii.gz

In this case, t2w (T2-weighted) is missing

Detecting missing modality during inference

When presenting three image files in each test folder, if you wish to automatically figure out which one is missing, we provide a script to do it. Please check detect_missing_modality.py.

Please note that after synthesizing the missing one, you DO NOT need to copy the other three files to the output folder. Please check the end of the Python script. Afterward, we will do automated segmentation based on the four modalities (three real images + one fake image).

Please, DO NOT add any suffixes, such as '-synthetic.nii.gz' to the output.

Performing segmentation using the synthetic image as a part of the input

After generating the missing image, segmentation can be performed by taking three real modalities + the missing modality as the input. We use the FeTS Consensus Models for image segmentation, the command is:

${fets_root_dir}/bin/FeTS_CLI_Segment -d /path/to/output/DataForFeTS \ # data directory after invoking ${fets_root_dir}/bin/PrepareDataset

  -a fets_singlet,fets_triplet \ # can be used with all pre-trained models currently available in FeTS

  -lF STAPLE,ITKVoting,SIMPLE,MajorityVoting \ # if a single architecture is used, this parameter is ignored

  -g 1 \ # '0': cpu, '1': request gpu

  -t 0 # '0': inference mode, '1': training mode

brasyn's Issues

Question regarding evaluation process

Firstly, thank you for providing the code related to the Brasyn challenge. While running the code, I have a question about measuring SSIM, specifically regarding the range of data values.

Should both array_real and array_syn, that is, GT and output, be in an unnormalized state? If not, the process of percentile clipping seems a bit unusual to me.

Additionally, I am curious about the reason why the prediction value is multiplied by two at line 66 in ssim.py.

Thank you .