DHT-Net:Dynamic Hierarchical Transformer Network for Liver and Tumor Segmentation

This is the official pytorch implementation of the DHT-Net:

Requirements

CUDA 11.0

PyTroch 1.7.0

Python 3.8

Torchvision 0.8.2

Usage

Installation

Install Pytorch1.7, nnUNet as below:

pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html

pip install nnunet

Data preprocessing

1.Download LiTS Dataset and 3DIRCADb Dataset

2.Preprocess the dataset according to the nnUNet.

• Dataset path

you should divide the training data,training labels,test data, test labels into four part, with the specific path (LiTS dataset as an example):

./LiTS/data_raw/DHTNet_raw_data/TaskXXX_LiTS/imagesTr

./LiTS/data_raw/DHTNet_raw_data/TaskXXX_LiTS/imagesTs

./LiTS/data_raw/DHTNet_raw_data/TaskXXX_LiTS/labelsTr

./LiTS/data_raw/DHTNet_raw_data/TaskXXX_LiTS/labelsTs

XXX is the inter identifier associated with your task name.

• cd DHTNet/dhtnet/path.py

you shuould set your dataset base path base, your preprocessing output path preprocessing_output_dir and your network training output dir network_training_output_dir_base, as example:

base = ./LiTS/data_raw/DHTNet_raw_data/

preprocessing_output_dir = ./LiTS/data_preprocessed/

network_training_output_dir_base = ./LiTS/Result/

you can also specify any data path in DHTNet/dhtnet/path.py.

3.Opertional data preprocessing.

• cd DHTNet/dhtnet/ (set the base_folder = base/TaskXXX_LiTS in gen_json.py)

Run Python gen_json.py

• cd DHTNet/dhtnet/experiment_planning/

Run Python nnUNet_plan_and_preprocess.py -t XXX

XXX is the inter identifier associated with your Task name.

Training

• cd DHTNet/dhtnet/training/

Python run_training -t XXX -f X

X specifies which fold of the 5-fold-cross-validation is trained.

Testing

Python run_training -t XXX -f X -val

Inference

• cd DHTNet/dhtnet/inference/

Python predict_simple.py -i INPUT_FOLDER -o OUTPUT_FOLDER -t xxx -f x -chk model_best

you can ensembling the predictions from several configurations with the following command:

Python ensenble_predictions.py -f FOLDER1 FOLDER2 ... -pp POSTPROCESSING_FILE

5. Acknowledgements

Part of codes are reused from the nnU-Net. Thanks to Fabian Isensee for the codes of nnU-Net.

Contact

Ruiyang Li(liruiyang@stu.xidian.edu.cn)