Awesome

[MICCAI2023] DHC

This repo is the official implementation of DHC: Dual-debiased Heterogeneous Co-training Framework for Class-imbalanced Semi-supervised Medical Image Segmentation which is accepted at MICCAI-2023.

🚀🚀🚀 We highly recommend you try our new work: https://github.com/xmed-lab/GenericSSL, which considers more practical scenarios of semi-supervised segmentation and the paper is accepted at NeurIPS-2023!

1. Environment

This code has been tested with Python 3.6, PyTorch 1.8, torchvision 0.9.0, and CUDA 11.1 on Ubuntu 20.04.

Before running the code, set the PYTHONPATH to pwd:

export PYTHONPATH=$(pwd)/code:$PYTHONPATH

2. Data Preparation

2.1 Synapse

The MR imaging scans are available at https://www.synapse.org/#!Synapse:syn3193805/wiki/. Please sign up and download the dataset.

Put the data in anywhere you want then change the file paths in config.py.

Run ./code/data/preprocess.py to

• convert .nii.gz files into .npy for faster loading.
• generate the train/validation/test splits
• generate the labeled/unlabeled splits

🔥🔥🔥 The preprocessed Synapse dataset is available for downloading via this link.

After preprocessing, the ./synapse_data/ folder should be organized as follows:

./synapse_data/
├── npy
│   ├── <id>_image.npy
│   ├── <id>_label.npy
├── splits
│   ├── labeled_20p.txt
│   ├── unlabeled_20p.txt
│   ├── train.txt
│   ├── eval.txt
│   ├── test.txt
│   ├── ...

2.2 AMOS

The dataset can be downloaded from https://amos22.grand-challenge.org/Dataset/

Run ./code/data/preprocess_amos.py to pre-process.

🔥🔥🔥 The preprocessed AMOS22 dataset is available for downloading via this link.

3. Training & Testing & Evaluating

Run the following commands for training, testing and evaluating.

bash train3times_seeds_20p.sh -c 0 -t synapse -m dhc -e '' -l 3e-2 -w 0.1

20p denotes training with 20% labeled data, you can change this to 2p, 5p, ... for 2%, 5%, ... labeled data.

Parameters:

-c: use which gpu to train

-t: task, can be synapse or amos

-m: method, dhc is our proposed method, other available methods including:

• cps
• uamt
• urpc
• ssnet
• dst
• depl
• adsh
• crest
• simis
• acisis
• cld

-e: name of current experiment

-l: learning rate

-w: weight of unsupervised loss

Weights of all the above models trained on 20% labeled Synapse can be downloaded from here.

Weights of all the above models trained on 5% labeled AMOS can be downloaded from here.

4. Results

4.1 Synapse

13 classes: Sp: spleen, RK: right kidney, LK: left kidney, Ga: gallbladder, Es: esophagus, Li: liver, St: stomach, Ao: aorta, IVC: inferior vena cava, PSV: portal & splenic veins, Pa: pancreas, RAG: right adrenal gland, LAG: left adrenal gland.

4.1.1 Trained with 10% labeled data

4.1.2 Trained with 20% labeled data

4.1.3 Trained with 40% labeled data

4.2 AMOS

15 classes: spleen, right kidney, left kidney, gallbladder, esophagus, liver, stomach, aorta, inferior vena cava, pancreas, right adrenal gland, left adrenal gland, duodenum, bladder, prostate/uterus

4.2.1 Trained with 2% labeled data

4.2.2 Trained with 5% labeled data

4.2.3 Trained with 10% labeled data

Cite

If this code is helpful for your study, please cite:

@inproceedings{wang2023dhc,
  title={DHC: Dual-debiased Heterogeneous Co-training Framework for Class-imbalanced Semi-supervised Medical Image Segmentation},
  author={Wang, Haonan and Li, Xiaomeng},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={582--591},
  year={2023},
  organization={Springer}
}

License

This repository is released under MIT License.