Mingen Zhang, Yuanyuan Gu, Meng Wang, Lei Mou, Jingfeng Zhang, Yitian Zhao
Medical Image Analysis (MedIA), 2026

YoloSeg is a two-stage framework for medical image segmentation using only one labeled image.

1. Foundation model-drvien pseudo-label generation
   Generating multi-view pseudo labels and divergence masks from a single labeled image using SAM2.

2. Robust pseudo-label learning for segmentation model
   Training a segmentation network with dual-component loss and cross-patch data augmentation.

We validated YoloSeg on 10 diverse public datasets, achieving an average Dice only 3.08% lower than the fully supervised baseline.

git clone https://github.com/iMED-Lab/YoloSeg.git
cd YoloSeg

conda create -n yoloseg python=3.10 -y
conda activate yoloseg

pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu118

pip install -e ./code_pl
pip install -r requirements.txt

Please organize the dataset as follows:

DatasetName/
├── file_list/
│   ├── train_all_frames.txt
│   ├── train_label_frames.txt
│   └── test_all_frames.txt
├── Train/
│   ├── JPEGImages/
│   │   ├── XXX001.png
│   │   ├── XXX002.png
│   │   └── ...
│   └── Annotations/
│       ├── XXX001.png
│       ├── XXX002.png
│       └── ...
└── Test/
    ├── JPEGImages/
    │   ├── XXX101.png
    │   ├── XXX102.png
    │   └── ...
    └── Annotations/
        ├── XXX101.png
        ├── XXX102.png
        └── ...

Notes:

• train_all_frames.txt: all training image filenames
• train_label_frames.txt: the selected labeled image filename for the one-shot setting
• test_all_frames.txt: all test image filenames
• Filenames in file_list/*.txt should be plain filenames, for example: XXX001.png
• Input images should be 3-channel .png
• Ground-truth labels should be single-channel .png
• Pseudo labels generated in Stage 1 are also single-channel .png

We provide a dataset structure example in: YoloSeg/data/ISIC2016

Please download SAM2 checkpoints from the official repository: SAM2 Official Repository

Then place the downloaded checkpoint files under: YoloSeg/code_pl/checkpoints/

For example:

YoloSeg/
└── code_pl/
    └── checkpoints/
        ├── sam2.1_hiera_tiny.pt
        ├── sam2.1_hiera_small.pt
        ├── sam2.1_hiera_base_plus.pt
        └── sam2.1_hiera_large.pt

We recommend using sam2.1_hiera_small.pt by default.

Run Stage 1 to generate:

• pl_original
• pl_rotate
• pl_flip
• divergence_mask

python code_pl/multi_view_inference.py \
  --data-root /path/to/DatasetName \
  --checkpoint code_pl/checkpoints/sam2.1_hiera_small.pt \
  --cfg code_pl/configs/sam2.1/sam2.1_hiera_s.yaml

After Stage 1, the Train/ directory will be automatically updated as:

Train/
├── JPEGImages/
├── Annotations/
├── pl_original/
├── pl_rotate/
├── pl_flip/
└── divergence_mask/

Train the segmentation model with the generated pseudo labels:

python code_seg/train.py \
  --data-root /path/to/DatasetName \
  --exp-name yoloseg_unet \
  --num-classes 2 \
  --image-size 256 \
  --batch-size 4 \
  --epochs 100

Run testing with the trained model:

python code_seg/test.py \
  --data-root /path/to/DatasetName \
  --checkpoint checkpoints/yoloseg_unet/best.pth \
  --output-dir outputs/yoloseg_unet_test \
  --num-classes 2 \
  --image-size 256

We would like to thank the authors of the following open-source projects:

• SAM2
• SSL4MIS

Their excellent work has greatly inspired and supported this project.

If you find YoloSeg useful, please cite:

@article{yoloseg2026,
  title   = {YoloSeg: You Only Label Once for Medical Image Segmentation},
  author  = {Zhang, Mingen and Gu, Yuanyuan and Wang, Meng and Mou, Lei and Zhang, Jingfeng and Zhao, Yitian},
  journal = {Medical Image Analysis},
  year    = {2026}
}

If you have any questions, feel free to contact: zhangmingen@nimte.ac.cn