First make sure you have all the proper paths set in paths.py: base, preprocessing_output_dir and network_training_output_dir.
Then copy the downloaded dataset into base/nnUNet_raw. For Task04_Hippocampus, for example, this should look like this:
base/nnUNet_raw/Task04_Hippocampus. Hereby, Task04_Hippocampus has three subfolders
(imagesTr, labelsTr, imagesTs) and a dataset.json file.
You can run the preprocessing and experiment planning for this stask by executing
python experiment_planning/plan_and_preprocess_task.py -t Task04_Hippocampus -p 8
For historical reasons nnU-Net does not like 4D niftis, so the first preprocessing step done by nnU-Net will be splitting the
4D niftis into a series of 3D niftis. They will be stored in base/nnUNet_raw_splitted when you run the preprocessing.
In this file we provide a description on how you need to convert your dataset to make it compatible with nnU-Net.
For the purpose of this
manual, we refer to your dataset as TaskXX_MY_DATASET. Hereby, XX is is a dual-digit number and MY_DATASET can
be anything you want.
We follow the folder structure of the Medical Segmentation Decathlon (MSD). In the splitted_4d_output_dir (see paths.py)
, create a subfolder
called TaskXX_MY_DATASET. In that subfolder, create the following three directories: imagesTr, labelsTr(, imagesTs).
These are for training images, training labels (and test images), respectively.
Just like the MSD we use the nifti (.nii.gz) file format. There is a crucial difference though.
While the MSD provides 4D niftis where the first axis is for the modality, we prefer to split each training case into
separate 3D Niftis. So what was a File patientID.nii.gz containing an image of shape (4, 160, 190, 160) now is four files with shape
(160, 190, 160) each. These four files should be named patientID_0000.nii.gz, patientID_0001.nii.gz, patientID_0002.nii.gz,
patientID_0003.nii.gz, where the ending 4-digit represents the modality. patientID can be anything you want. Make
sure that the same modality is assigned the same digit for all patients (for example if you have T1 and T2 then you
need to make sure T1 is always 0000 and T2 always 0001). If you data has only one modality, just name your cases patientID_0000.nii.gz.
Copy all training cases (just the data, not the labels) into the imagesTr subfolder.
Copy training labels into the labelsTr subfolder. Labels are always 3D niftis and should be named patientID.nii.gz
where patientiID is identical to the identifier used for the corresponding raw data.
If you have test images, convert them to 3D Nifti like the training data and place them in the imagesTs folder.
All data you wish to predict with nnU-Net must be in this format.
Finally you need to create a dataset.json file that you place in the TaskXX_MY_DATASET root folder. This file was
always provided in the MSD challenge and is therefore a requirement for nnU-Net. Have a look at (for example)
LiverTumorSegmentationChallenge.py to see what it needs to look like. Important: The list stored in 'training' contains images and labels are
both called patientID.nii.gz (dropping the 0000 ending we used here) for consistency!