nnU-Net requires the raw data to be brought into a specific format so that it know how to read and interpret it. This format closely, but not entirely, follows the format used by the Medical Segmentation Decathlon (MSD).
The entry point to nnU-Net is the nnUNet_raw_data_base folder (which the user needs to specify when installing nnU-Net!). Each segmentation dataset is stored as a separate 'Task'. Tasks are associated with a task ID, a three digit integer (this is different from the MSD!) and a task name (which you can freely choose): Task005_Prostate has 'Prostate' as task name and the task id is 5. Tasks are stored in the nnUNet_raw_data_base/nnUNet_raw_data folder like this:
nnUNet_raw_data_base/nnUNet_raw_data/
├── Task001_BrainTumour
├── Task002_Heart
├── Task003_Liver
├── Task004_Hippocampus
├── Task005_Prostate
├── ...
Within each task folder, the following structure is expected:
Task001_BrainTumour/
├── dataset.json
├── imagesTr
├── (imagesTs)
└── labelsTr
Please make your custom task ids start at 500 to ensure that there will be no conflicts with downloaded pretrained models!!! (IDs cannot exceed 999)
imagesTr contains the images belonging to the training cases. nnU-Net will run pipeline configuration, training with cross-validation, as well as finding postprocesing and the best ensemble on this data. imagesTs (optional) contains the images that belong to the test cases , labelsTr the images with the ground truth segmentation maps for the training cases. dataset.json contains metadata of the dataset.
Each training case is associated with an identifier = a unique name for that case. This identifier is used by nnU-Net to recognize which label file belongs to which image. All images (including labels) must be 3D nifti files (.nii.gz)!
The image files can have any scalar pixel type. The label files must contain segmentation maps that contain consecutive integers, starting with 0: 0, 1, 2, 3, ... num_labels. 0 is considered background. Each class then has its own associated integer value. Images may have multiple modalities. This is especially often the case for medical images. Modalities are very much like color channels in photos (three color channels: red, green blue), but can be much more diverse: CT, different types or MRI, and many other. Imaging modalities are identified by nnU-Net by their suffix: a four-digit integer at the end of the filename. Imaging files must therefore follow the following naming convention: case_identifier_XXXX.nii.gz. Hereby, XXXX is the modality identifier. What modalities these identifiers belong to is specified in the dataset.json file (see below). Label files are saved as case_identifier.nii.gz
This naming scheme results in the following folder structure. It is the responsibility of the user to bring their data into this format!
Here is an example for the first Task of the MSD: BrainTumour. Each image has four modalities: FLAIR (0000), T1w (0001), T1gd (0002) and T2w (0003). Note that the imagesTs folder is optional and does not have to be present.
nnUNet_raw_data_base/nnUNet_raw_data/Task001_BrainTumour/
├── dataset.json
├── imagesTr
│ ├── BRATS_001_0000.nii.gz
│ ├── BRATS_001_0001.nii.gz
│ ├── BRATS_001_0002.nii.gz
│ ├── BRATS_001_0003.nii.gz
│ ├── BRATS_002_0000.nii.gz
│ ├── BRATS_002_0001.nii.gz
│ ├── BRATS_002_0002.nii.gz
│ ├── BRATS_002_0003.nii.gz
│ ├── BRATS_003_0000.nii.gz
│ ├── BRATS_003_0001.nii.gz
│ ├── BRATS_003_0002.nii.gz
│ ├── BRATS_003_0003.nii.gz
│ ├── BRATS_004_0000.nii.gz
│ ├── BRATS_004_0001.nii.gz
│ ├── BRATS_004_0002.nii.gz
│ ├── BRATS_004_0003.nii.gz
│ ├── ...
├── imagesTs
│ ├── BRATS_485_0000.nii.gz
│ ├── BRATS_485_0001.nii.gz
│ ├── BRATS_485_0002.nii.gz
│ ├── BRATS_485_0003.nii.gz
│ ├── BRATS_486_0000.nii.gz
│ ├── BRATS_486_0001.nii.gz
│ ├── BRATS_486_0002.nii.gz
│ ├── BRATS_486_0003.nii.gz
│ ├── BRATS_487_0000.nii.gz
│ ├── BRATS_487_0001.nii.gz
│ ├── BRATS_487_0002.nii.gz
│ ├── BRATS_487_0003.nii.gz
│ ├── BRATS_488_0000.nii.gz
│ ├── BRATS_488_0001.nii.gz
│ ├── BRATS_488_0002.nii.gz
│ ├── BRATS_488_0003.nii.gz
│ ├── BRATS_489_0000.nii.gz
│ ├── BRATS_489_0001.nii.gz
│ ├── BRATS_489_0002.nii.gz
│ ├── BRATS_489_0003.nii.gz
│ ├── ...
└── labelsTr
├── BRATS_001.nii.gz
├── BRATS_002.nii.gz
├── BRATS_003.nii.gz
├── BRATS_004.nii.gz
├── ...
Here is another example of the second task of the MSD, which has only one modality:
nnUNet_raw_data_base/nnUNet_raw_data/Task002_Heart/
├── dataset.json
├── imagesTr
│ ├── la_003_0000.nii.gz
│ ├── la_004_0000.nii.gz
│ ├── ...
├── imagesTs
│ ├── la_001_0000.nii.gz
│ ├── la_002_0000.nii.gz
│ ├── ...
└── labelsTr
├── la_003.nii.gz
├── la_004.nii.gz
├── ...
For each training case, all images must have the same geometry to ensure that their pixel arrays are aligned. Also make sure that all your data is co-registered!
The dataset.json file used by nnU-Net is identical to the ones used by the MSD. For your custom tasks you need to create them as well and thereby exactly follow the same structure. This is where you can download the MSD data for reference.
NEW: There now is a utility with which you can generate the dataset.json automatically. You can find it
here (look for the function generate_dataset_json).
See Task120 for an example on how to use it. And read
its documentation!
Here is the content of the dataset.json from the Prostate task:
{
"name": "PROSTATE",
"description": "Prostate transitional zone and peripheral zone segmentation",
"reference": "Radboud University, Nijmegen Medical Centre",
"licence":"CC-BY-SA 4.0",
"relase":"1.0 04/05/2018",
"tensorImageSize": "4D",
"modality": {
"0": "T2",
"1": "ADC"
},
"labels": {
"0": "background",
"1": "PZ",
"2": "TZ"
},
"numTraining": 32,
"numTest": 16,
"training":[{"image":"./imagesTr/prostate_16.nii.gz","label":"./labelsTr/prostate_16.nii.gz"},{"image":"./imagesTr/prostate_04.nii.gz","label":"./labelsTr/prostate_04.nii.gz"},...],
"test": ["./imagesTs/prostate_08.nii.gz","./imagesTs/prostate_22.nii.gz","./imagesTs/prostate_30.nii.gz",...]
}
Note that we truncated the "training" and "test" lists for clarity. You need to specify all the cases in there. If you
don't have test images (imagesTs does not exist) you can leave "test" blank: "test": [].
Please also have a look at the python files located here. They show how we created our custom dataset.jsons for a range of public datasets.
The previous release of nnU-Net allowed users to either start with 4D or 3D niftis. This resulted in some confusion, however, because some users would not know where they should save their data. We therefore dropped support for the 4D niftis used by the MSD. Instead, we provide a utility that converts the MSD datasets into the format specified above:
nnUNet_convert_decathlon_task -i FOLDER_TO_TASK_AS_DOWNLOADED_FROM_MSD -p NUM_PROCESSESFOLDER_TO_TASK_AS_DOWNLOADED_FROM_MSD needs to point to the downloaded task folder (such as Task05_Prostate, note the
2-digit task id!). The converted Task will be saved under the same name in nnUNet_raw_data_base/nnUNet_raw_data
(but with a 3 digit identifier). You can overwrite the task id of the converted task by using the -output_task_id option.
nnU-Net was originally built for 3D images. It is also strongest when applied to 3D segmentation problems because a large proportion of its design choices were built with 3D in mind. Also note that many 2D segmentation problems, especially in the non-biomedical domain, may benefit from pretrained network architectures which nnU-Net does not support. Still, there is certainly a need for an out of the box segmentation solution for 2D segmentation problems. And also on 2D segmentation tasks nnU-Net cam perform extremely well! We have, for example, won a 2D task in the cell tracking challenge with nnU-Net (see our Nature Methods paper) and we have also successfully applied nnU-Net to histopathological segmentation problems.
Working with 2D data in nnU-Net requires a small workaround in the creation of the dataset. Essentially, all images must be converted to pseudo 3D images (so an image with shape (X, Y). When working with 2D images it is important to follow the correct axis ordering. When loading the images with SimpleITK, the resulting numpy array shape should be (1, x, y). We recommend you save your images with SimpleITK so that the correct shape is guaranteed. If you prefer to save your images with nibabel, please save them as (y, x, 1) (SimpleITK reverts the ordering of the axes when reading). To check whether your 2D images have the correct shape you can run the following snippet:
import SimpleITK as sitk
print(sitk.GetArrayFromImage(sitk.ReadImage(FILENAME)).shape)
The resulting image must be saved in nifti format. Hereby it is important to set the spacing of the first axis (the one with shape 1) to a value larger than the others. If you are working with niftis anyways, then doing this should be easy for you. This example here is intended for demonstrating how nnU-Net can be used with 'regular' 2D images. We selected the massachusetts road segmentation dataset for this because it can be obtained easily, it comes with a good amount of training cases but is still not too large to be difficult to handle.
See here for an example. This script contains a lot of comments and useful information. Also have a look here.
When updating a dataset you not only need to change the data located in nnUNet_raw_data_base/nnUNet_raw_data. Make
sure to also delete the whole (!) corresponding dataset in nnUNet_raw_data_base/nnUNet_cropped_data. nnU-Net will not
repeat the cropping (and thus will not update your dataset) if the old files are still in nnUNet_cropped_data!
The best way of updating an existing dataset is (choose one):
- delete all data and models belonging to the old version of the dataset (nnUNet_preprocessed, corresponding results in RESULTS_FOLDER/nnUNet, nnUNet_cropped_data, nnUNet_raw_data), then update
- (recommended) create the updated dataset from scratch using a new task ID and name
Please have a look at the following tasks: