Official code for "A Semantic Knowledge Complementarity based Decoupling Framework for Semi-supervised Class-imbalanced Medical Image Segmentation". (CVPR 2025)

• The Synapse dataset can be downloaded from https://www.synapse.org/#!Synapse:syn3193805/wiki/.

  Run preprocess.py to convert .nii.gz files into .npy.

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

  Run preprocess_amos.py to convert .nii.gz files into .npy.

• The format of the preprocessed data is .npy with a size of 80×160×160. We will upload the preprocessed data as soon as possible.

  The preprocessed Synapse dataset has been uploaded to https://pan.baidu.com/s/1TO9-PE4ZkjPMdw_QPIVSPQ?pwd=zae8

• The data splits have been uploaded to the codebase.

When training on the Synapse dataset, we conducted 3 experiments with random seeds of 0,1,666 respectively. The other hyperparameters are as follows:

max_epoch=1500, cps_loss='w_ce+dice', sup_loss='w_ce+dice', batch_size=2, num_workers=2, base_lr=0.3, ema_w=0.99, cps_w=10, cps_rampup=True, consistency_rampup=None

When training on the AMOS dataset, we conducted 1 experiment with random seeds of 0. The other hyperparameters are as follows:

max_epoch=1500, cps_loss='w_ce+dice', sup_loss='w_ce+dice', batch_size=2, num_workers=2, base_lr=0.1, ema_w=0.99, cps_w=10, cps_rampup=True, consistency_rampup=None

Then run train_skcdf.py to train.

Run test.py to generate prediction results.

Run evaluate_Ntimes.py to calculate average Dice and ASD.

If this code is useful for your research, please cite:

@inproceedings{zhang2025semantic,
  title={A Semantic Knowledge Complementarity based Decoupling Framework for Semi-supervised Class-imbalanced Medical Image Segmentation},
  author={Zhang, Zheng and Yin, Guanchun and Zhang, Bo and Liu, Wu and Zhou, Xiuzhuang and Wang, Wendong},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={25940--25949},
  year={2025}
}

This project is built upon the following outstanding open-source projects: GenericSSL, AllSpark and ABC. We deeply appreciate their contributions to the community.