farrell236/AscitesModel

Ascites Segmentation with nnUNet

This model was trained as part of the research 'Deep Learning Segmentation of Ascites on Abdominal CT Scans for Automatic Volume Quantification' (Paper, arXiv).

Method 1: Run Inference using nnunet_predict.py

1. Install nnUNet_v1 and PyTorch.

user@machine:~/ascites_segmentation$ pip install torch torchvision torchaudio nnunet matplotlib

2. Run inference with command:

user@machine:~/ascites_segmentation$ python nnunet_predict.py -i file_list.txt -t TMP_DIR -o OUTPUT_FOLDER -m /path/to/nnunet/model_weights

usage: tmp.py [-h] [-i INPUT_LIST] -t TMP_FOLDER -o OUTPUT_FOLDER -m MODEL [-v]

Inference using nnU-Net predict_from_folder Python API

optional arguments:
  -h, --help            show this help message and exit
  -i INPUT_LIST, --input_list INPUT_LIST
                        Input image file_list.txt
  -t TMP_FOLDER, --tmp_folder TMP_FOLDER
                        Temporary folder
  -o OUTPUT_FOLDER, --output_folder OUTPUT_FOLDER
                        Output Segmentation folder
  -m MODEL, --model MODEL
                        Trained Model
  -v, --verbose         Verbose Output

N.B.

• model_weights folder should contain fold0, fold1, etc...
• WARNING: the program will try to create file links first, but will fallback to filecopy if fails

Method 2: Run Inference using nnUNet_predict from shell

1. Install nnUNet_v1 and PyTorch.

user@machine:~/ascites_segmentation$ pip install torch torchvision torchaudio nnunet matplotlib

2. Setup environment variables so that nnU-Net knows where to find trained models:

user@machine:~/ascites_segmentation$ export nnUNet_raw_data_base="/absolute/path/to/nnUNet_raw_data_base"
user@machine:~/ascites_segmentation$ export nnUNet_preprocessed="/absolute/path/to/nnUNet_preprocessed"
user@machine:~/ascites_segmentation$ export RESULTS_FOLDER="/absolute/path/to/nnUNet_trained_models"

3. Run inference with command:

user@machine:~/ascites_segmentation$ nnUNet_predict -i INPUT_FOLDER -o OUTPUT_FOLDER -t 505 -m 3d_fullres -f N --save_npz 

where:

• -i: input folder of .nii.gz scans to predict. NB, filename needs to end with _0000.nii.gz to tell nnU-Net only one kind of modality
• -o: output folder to store predicted segmentations, automatically created if not exist
• -t 505: (do not change) Ascites pretrained model name
• -m 3d_fullres (do not change) Ascites pretrained model name
• N: Ascites pretrained model fold, can be [0, 1, 2, 3, 4]
• --save_npz: save softmax scores, required for ensembling multiple folds

Optional [Additional] Inference Steps

a. use nnUNet_find_best_configuration to automatically get the inference commands needed to run the trained model on data.

b. ensemble predictions using nnUNet_ensemble by running:

user@machine:~/ascites_segmentation$ nnUNet_ensemble -f FOLDER1 FOLDER2 ... -o OUTPUT_FOLDER -pp POSTPROCESSING_FILE

where FOLDER1 and FOLDER2 are predicted outputs by nnUNet (requires --save_npz when running nnUNet_predict).

Method 3: Docker Inference

Requires nvidia-docker to be installed on the system (Installation Guide). This nnunet_docker predicts ascites with all 5 trained folds and ensembles output to a single prediction.

1. Build the nnunet_docker image from Dockerfile:

user@machine:~/ascites_segmentation$ sudo docker build -t nnunet_docker .

2. Run docker image on test volumes:

user@machine:~/ascites_segmentation$ sudo docker run \
--gpus 0 \
--volume /absolute/path/to/INPUT_FOLDER:/tmp/INPUT_FOLDER \
--volume /absolute/path/to/OUTPUT_FOLDER:/tmp/OUTPUT_FOLDER \
nnunet_docker /bin/sh inference.sh

• --gpus parameter:

  • 0, 1, 2, ..., n for integer number of GPUs
  • all for all available GPUs on the system
  • '"device=2,3"' for specific GPU with ID

• --volume parameter

  • /absolute/path/to/INPUT_FOLDER and /absolute/path/to/OUTPUT_FOLDER folders on the host system needs to be specified
  • INPUT_FOLDER contains all .nii.gz volumes to be predicted
  • predicted results will be written to OUTPUT_FOLDER

Citation

If you find this repository helpful in your research, please consider citing our paper:

@article{hou2024deep,
  title={Deep Learning Segmentation of Ascites on Abdominal CT Scans for Automatic Volume Quantification},
  author={Hou, Benjamin and Lee, Sung-Won and Lee, Jung-Min and Koh, Christopher and Xiao, Jing and Pickhardt, Perry J. and Summers, Ronald M.}
  journal={Radiology: Artificial Intelligence},
  pages={e230601},
  year={2024},
  publisher={Radiological Society of North America}
}