RPR-Loc

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

Data format

the data format should be like:

├── train
│   └── 001
│         └── data.nii.gz
|   └── 002
│         └── data.nii.gz
|   .
|   .
|   .
└── valid
|   └── 00N
|      └── data.nii.gz
|   .
|   .
|   .
└── test
|   └── 00X
|      └── data.nii.gz
|   .
|   .
|   .

The name of data division folder should be train/valid/test while the folder name of each volume could be arbitrary.

Preprocess

The preprocess contains two stage:

1. Resample: use data_process/Resample_data.py
2. Normalization: use data_process/Normalize_data.py

Training

Use train/train_position.py.
The config file is in config, containing 4 files for coarse/fine & pancreas/head and neck dataset.
You can choose anyone by changing the config_file in line 173 of train.py.