RPR-Loc

Official code for MICCAI2021. 'Contrastive Learning of Relative Position Regression for One-Shot Object Localization in 3D Medical Images'

Dataset

You could download the processed dataset from: StructSeg task1 (Organ-at-risk segmentation from head & neck CT scans): BaiDu Yun or Google Drive and TCIA-Pancreas: BaiDu Yun or Google Drive into data/ and unzip them. For TCIA-Pancreas, please cite the original paper (Deeporgan: Multi-level deep convolutional networks for automated pancreas segmentation).

Training

The training config is in config/train/, containing 4 files for coarse/fine & pancreas/head and neck dataset. You could change the parameters in configuration txt for your own experiments. More details could be found in config/train/readme.md.

For example, you could cd train then run python train_position.py ../config/train/train_position_pancreas_coarse.txt to train a coarse RPR model for TCIA-Pancreas dataset.

Pre-trained models

We provide the pre-trained Pnet_2 model for TCIA-Pancreas dataset. You could download the coarse model from Google Drive and the fine model from Google Drive. Both two models shoud be saved in weights/Pnet_2/.

Detection

For example, you could cd detection then run python c2f_detection.py ../config/test/test_c2f_pancreas_detection.txt for pancreas detection (the default setting requires the pre-trained models). You could change the parameters in configuration txt for your own experiments. More details could be found in config/test/readme.md.

Performance

We have fixed a bug in our original framework and increase the detection IOU of pancreas from 0.495 in the paper to 0.58.

Citation

If you find this research useful, please consider citing our work:

@inproceedings{lei2021contrastive,
title={Contrastive learning of relative position regression for one-shot object localization in 3D medical images},
author={Lei, Wenhui and Xu, Wei and Gu, Ran and Fu, Hao and Zhang, Shaoting and Zhang, Shichuan and Wang, Guotai},
booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
pages={155--165},
year={2021},
organization={Springer}
}