CoDG-Net : Structure-Guided Style Diffusion and Collaborative Learning to Mitigate Catastrophic Forgetting in Medical Image Domain Generalization

[Paper]  |  [arXiv]  |  [Project Page]  |  [Poster]  |  [BibTeX]

Links will be updated upon paper publication.

---

Introduction

Abstract: Domain generalization (DG) for medical image segmentation is both highly challenging and critically important. However, existing medical DG methods largely overlook the issue of catastrophic forgetting (CF). When pursuing cross-domain robustness, models often sacrifice their retention of source-domain knowledge. This can directly threaten diagnostic safety in already-deployed clinical scenarios. To address this, we investigate data augmentation strategies and catastrophic forgetting for medical image DG segmentation. First, we propose a structure-guided style diffusion augmentation method. Constrained by anatomical structure consistency in the frequency domain, this method performs cross-domain diffusion and recombination on the amplitude spectrum, generating pseudo-domain samples with more diverse and broader style coverage to better support domain generalization. Then, we design a collaborative learning network with a dual-branch interactive architecture (CoDG-Net), together with a novel learning bias-guided strategy that adaptively regulates knowledge transfer at both the layer level and the task level, thereby effectively mitigating catastrophic forgetting on the source domain. Experiments and ablation studies on single-source and multi-source medical DG benchmark datasets demonstrate that CoDG-Net not only outperforms existing state-of-the-art methods in target-domain segmentation performance, but also achieves a lower forgetting rate on the source-domain data.

Repository Structure

CoDG-Net/
├── dataloaders/              
├── datasets/              
│   ├── BraTS2018/
│   └── Fundus/
├── images/                  
├── models/                  
├── preprocess/       
│   ├── bezier_curve.py
│   ├── split_data.py
│   └── preprocess_func.py
├── train-sh/                
│   ├── Diff-train/
│   │   ├── brats_2018/
│   │   └── fundus/
│   └── CoDG-train/
│       ├── brats_2018/
│       └── fundus/
├── test-sh/                  
│   ├── Diff-test/
│   │   ├── brats_2018/
│   │   └── fundus/
│   └── CoDG-test/
│       ├── brats_2018/
│       └── fundus/
├── utils/                    
├── train_diff.py             
├── test_diff.py              
├── train_codg.py              
├── test_codg.py               
├── LICENSE                  
└── README.md

Usage

1. First, download the dataset and place it in the 'datasets' folder

2. Preprocess the dataset with the function in preprocess/split_data.py and preprocess/preprocess_func.py

3. Train the Diffusion model with the training script:

   cd train-sh/Diff-train/brats_2018
   sh ./train-t2-1.sh

4. Test the Diffusion model with script:

   cd test-sh/Diff-test/brats_2018
   sh ./test-t2-1.sh

5. Train the CoDG-Net model with script:

   cd train-sh/CoDG-train/brats_2018
   sh ./train-t2-1.sh

   The CoDG-Net as follows:

   

6. Test the CoDG-Net model with script:

   cd test-sh/CoDG-test/brats_2018
   sh ./test-t2-1.sh

Citation

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

@inproceedings{xxx2026codgnet,
  title     = {CoDG-Net: Structure-Guided Style Diffusion and Collaborative Learning to Mitigate Catastrophic Forgetting in Medical Image Domain Generalization},
  author    = {xxx and xxx and xxx and xxx and xxx},
  booktitle = {Proceedings of the 35th International Joint Conference on Artificial Intelligence (IJCAI)},
  year      = {2026},
  pages     = {xxx--xxx},
  address   = {xxx},
}

The official BibTeX entry will be updated once the IJCAI 2026 proceedings are released.

Acknowledgement

The datasets used in this paper are downloaded from BraTS2018 and Fundus.