RFNet for Incomplete/Missing Multi-modal Brain Tumor Segmentation

Official implementation of RFNet: Region-aware Fusion Network for Incomplete Multi-modal Brain Tumor Segmentation), ICCV2021.

Results

Brats2020

All missing and full-set situations (15 situations) are considered during testing. The average results are reported here. Please refer to our paper for more details.

Method	Complete	Core	Enhancing
HeMIS	75.10	65.45	47.73
U-HVED	81.24	67.19	48.55
RobustSeg	84.17	73.45	55.49
RFNet (Ours)	86.98	78.23	61.47

Complete, Core and Enhancing denote the dice score (%) of the whole tumor, the tumor core and the enhancing tumor, respectively.

Brats2018

Brats2018 contains three different training and test splits and the average results are reported here.

Method	Complete	Core	Enhancing
HeMIS	78.60	59.70	48.10
U-HVED	80.10	64.00	50.00
RobustSeg	84.37	69.78	51.02
RFNet (Ours)	85.67	76.53	57.12

Brats2015

Method	Complete	Core	Enhancing
HeMIS	68.22	54.07	43.86
U-HVED	81.57	64.68	56.76
RobustSeg	84.45	69.19	57.33
RFNet (Ours)	86.13	71.93	64.13

Checkpoints and logs

Brats2020	Brats2018 split1	Brats2018 split2	Brats2018 split3	Brats2015
model	model	model	model	model
log	log	log	log	log

Installation

We use pytorch1.2.0 and cuda9.0.

For all datasets, we train our networks with 2 * V100 (16G).

get dataset and environment here and unzip them.

tar -xzf BRATS2020_Training_none_npy.tar.gz
tar -xzf BRATS2018_Training_none_npy.tar.gz
tar -xzf BRATS2015_Training_none_npy.tar.gz
tar -xzf pytorch_1.2.0a0+8554416-py36tf.tar.gz
tar -xzf cuda-9.0.tar.gz

Usage

1. Set dataname, pypath and cudapath in job.sh.

2. Set different splits for Brats2018 in L99-100 in train.py.

3. Then run:

bash job.sh

Note

1. We obtain the results by evaluating our models in the last epoch with the test set. If you want to evaluate models in other epochs, please use the --resume as in job.sh.

2. We also provide the preprocessing code preprocess.py. When using preprocess.py, you need to set the path of raw data 'src_path' and the path of processed data 'tar_path' in preprocess.py. The data structure in 'src_path' is shown as below:

BraTS20_Training_001/
    BraTS20_Training_001_flair.nii.gz
    BraTS20_Training_001_t1ce.nii.gz
    BraTS20_Training_001_t2.nii.gz
    BraTS20_Training_001_seg.nii.gz
    BraTS20_Training_001_t1.nii.gz
BraTS20_Training_002/
    BraTS20_Training_002_flair.nii.gz
    BraTS20_Training_002_t1ce.nii.gz
    BraTS20_Training_002_t2.nii.gz
    BraTS20_Training_002_seg.nii.gz
    BraTS20_Training_002_t1.nii.gz
BraTS20_Training_003/
...
...
BraTS20_Training_369/

Citation

@inproceedings{ding2021rfnet,
  title={RFNet: Region-Aware Fusion Network for Incomplete Multi-Modal Brain Tumor Segmentation},
  author={Ding, Yuhang and Yu, Xin and Yang, Yi},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={3975--3984},
  year={2021}
}