Given that nnUNet is a relatively flexible framework, we have developed a container that allows users to run nnUNet in a container while varying the necessary models. The main features are inferring all necessary parameters from the nnUNet files (spacing, extensions) and working for both DICOM folder and SITK-readable files. If the input is a DICOM, the segmentation is converted into a DICOM-seg file, compatible with PACS systems.
A considerable objective of this framework was its deployment as a standalone tool (for bash
). To use it:
- Install the necessary packages using an appropriate Python environment (i.e.
pip install -r requirements.txt
). We have tested this using Pythonv3.11
- Run
python utils/entrypoints.py --help
to see the available options - Segment away!
python utils/entrypoint.py --help
usage: Entrypoint for nnUNet prediction. Handles all data format conversions. [-h] --series_paths SERIES_PATHS [SERIES_PATHS ...] --model_path
MODEL_PATH [--checkpoint_name CHECKPOINT_NAME] --output_dir
OUTPUT_DIR --metadata_path METADATA_PATH [--study_uid STUDY_UID]
[--folds FOLDS [FOLDS ...]] [--tta] [--tmp_dir TMP_DIR] [--is_dicom]
[--proba_map] [--rt_struct_output] [--save_nifti_inputs]
options:
-h, --help show this help message and exit
--series_paths SERIES_PATHS [SERIES_PATHS ...], -i SERIES_PATHS [SERIES_PATHS ...]
Path to input series
--model_path MODEL_PATH, -m MODEL_PATH
Path to nnUNet model folder
--checkpoint_name CHECKPOINT_NAME, -ckpt CHECKPOINT_NAME
Checkpoint name for nnUNet
--output_dir OUTPUT_DIR, -o OUTPUT_DIR
Path to output directory
--metadata_path METADATA_PATH, -M METADATA_PATH
Path to metadata template for DICOM-Seg output
--study_uid STUDY_UID, -s STUDY_UID
Study UID if series are SimpleITK-readable files
--folds FOLDS [FOLDS ...], -f FOLDS [FOLDS ...]
Sets which folds should be used with nnUNet
--tta, -t Uses test-time augmentation during prediction
--tmp_dir TMP_DIR Temporary directory
--is_dicom, -D Assumes input is DICOM (and also converts to DICOM seg; prediction.dcm in output_dir)
--proba_map, -p Produces a Nifti format probability map (probabilities.nii.gz in output_dir)
--rt_struct_output Produces a DICOM RT Struct file (struct.dcm in output_dir)
--save_nifti_inputs, -S
Moves Nifti inputs to output folder (volume_XXXX.nii.gz in output_dir)
Example:
python utils/entrypoints.py \
-i study/series_1 study/series_2 study/series_3 \
-o example_output/ \
-m models/prostate_model \
-M metadata_templates/metadata-template.json \
-D -f 0 1 2 3 4 \
--proba_map \
--save_nifti_inputs
Firstly, users must install Docker. Docker requires sudo
access so users should be sure to have this. Then:
- Build the container (
sudo docker build -f Dockerfile . -t nnunet_predict
) - Run the container. We have replicated this as an additional script (
utils/entrypoint-with-docker.py
) with the same arguments as those specified to run as a standalone tool with the addition of a-c
flag specifying the name of the Docker image.
With utils/entrypoint-with-docker.py
, this:
docker run \
--gpus all \
--user "$(id -u):$(id -g)" \
-v $(dirname $(realpath $INPUT_PATHS)):/data/input \
-v $(realpath $OUTPUT_FOLDER):/data/output \
-v $(realpath $MODEL_FOLDER):/model \
-v $(dirname $(realpath $METADATA_TEMPLATE)):/metadata \
--rm \
$DOCKER_IMAGE \
-i $file_names_in_docker -d -M $metadata_name_in_docker
becomes this (for a DICOM input):
python utils/entrypoint-with-docker.py \
-i $INPUT_PATHS \
-o $OUTPUT_FOLDER \
-m $MODEL_FOLDER \
-d \
-M $METADATA_TEMPLATE \
-c $DOCKER_IMAGE
It is necessary to generate metadata templates for the conversion between the segmentation prediction volume and DICOM volumes. To generate these, the pydicom_seg
developers recommend this web app. It is easy to use and generates reliable metadata templates. Metadata templates should be generated for all segmentation targets to ensure that everything is correctly formatted.