We convert the training data list into json file using create_json.py.
The training and validation data are saved as follow:
train/
├── images
├── labels
validation/
├── images
├── labels
We specifiy the model hyperparmeters in the yaml files, sample yaml files for different model scales can be found in yaml folder.
yaml/
├── unest_base
├── unest_large
├── unest_small
Before training, specify the paths in the yaml files and set the yaml path in the main.py.
We use inference.py to test each images in the testing set and save the probability of each fold into .npy format for further ensemble. We use ensemble.py to ensemble the predictions from different folds and save the final prediction into Nifti format. The trained model weight can found here: renalseg_weight.
The results of different folds is saved as below:
pred_0.7/
├── fold0
├── fold1
├── fold2
├── fold3
├── fold4
Running inference:
python inference.py --imagesTs_path test_images_path --saved_model_path path2saved_model --base_dir output_path --fold 0 --overlap 0.7 --device 0
Running ensemble:
python ensemble.py --prob_dir output_from_inference --img_path test_images_path --out_path output_path
For developing publicly available segmentation tools, we introduce the MONAI Bundle module that supports building Python-based workflows via structured configurations. Renal Substructure Seg MONAI Boundle.
If you find this repository useful, please consider citing the following papers:
@article{yu2023unest,
title={UNesT: local spatial representation learning with hierarchical transformer for efficient medical segmentation},
author={Yu, Xin and Yang, Qi and Zhou, Yinchi and Cai, Leon Y and Gao, Riqiang and Lee, Ho Hin and Li, Thomas and Bao, Shunxing and Xu, Zhoubing and Lasko, Thomas A and others},
journal={Medical Image Analysis},
pages={102939},
year={2023},
publisher={Elsevier}
}