TMI 2025 | Co-Seg++: Mutual Prompt-Guided Collaborative Learning for Versatile Medical Segmentation

[arXiv]

Qing Xu¹ Yuxiang Luo² Wenting Duan³ Zhen Chen^4✉

¹University of Nottingham   ²Waseda University   ³Univeristy of Lincoln   ⁴Yale University  

^✉ Corresponding Author.

---

📰News

• [2025.12.05] Our Co-Seg++ has been accepted by IEEE Transactions on Medical Imaging (TMI) !
• [2025.06.03] We have released the code for Co-Seg++ !

🛠Setup

git clone https://github.com/xq141839/Co-Seg-Plus.git
cd Co-Seg-Plus
conda create -f Co-Seg-Plus.yaml

Key requirements: Cuda 12.2+, PyTorch 2.4+, mamba-ssm 2.1.0+

📚Data Preparation

• PUMA: Challenge Link

The data structure is as follows.

Co-Seg-Plus
├── datasets
│   ├── image_1024
│     ├── training_set_metastatic_roi_001.png
|     ├── ...
|   ├── mask_sem_1024
│     ├── training_set_metastatic_roi_001_nuclei.npy
|     ├── ...
|   ├── mask_ins_1024
│     ├── training_set_metastatic_roi_001_tissue.npy
|     ├── ...
|   ├── data_split.json

The json structure is as follows.

{ 
 "train": ['training_set_metastatic_roi_061.png'],
 "valid": ['training_set_metastatic_roi_002.png'],
 "test":  ['training_set_metastatic_roi_009.png'] 
 }

🎪Quickstart

• Train the Co-Seg++ with the default settings:

python train.py --dataset data/$YOUR DATASET NAME$ --sam_pretrain pretrain/$SAM2 CHECKPOINT$

📜Citation

If you find this work helpful for your project, please consider citing the following paper:

@ARTICLE{11299102,
  author={Xu, Qing and Luo, Yuxiang and Duan, Wenting and Chen, Zhen},
  journal={IEEE Transactions on Medical Imaging}, 
  title={Co-Seg++: Mutual Prompt-Guided Collaborative Learning for Versatile Medical Segmentation}, 
  year={2026},
  volume={45},
  number={5},
  pages={1947-1959}

Acknowledgements

• SAM2
• Medical-SAM-Adapter
• MambaVision
• CellViT