DINO-AugSeg

DINO-AugSeg — Exploiting DINOv3-Based Self-Supervised Features for Robust Few-Shot Medical Image Segmentation

📄 Paper:
Exploiting DINOv3-Based Self-Supervised Features for Robust Few-Shot Medical Image Segmentation

🌐 Hugging Face Paper Page: Link
🤖 Model: DINO-AugSeg
📂 Dataset: ACDC

---

📌 Overview

This repository provides the official implementation of DINO-AugSeg, a few-shot medical image segmentation framework that leverages pretrained DINOv3 self-supervised visual representations. By integrating feature-level augmentation and contextual fusion, DINO-AugSeg achieves strong performance under limited annotation settings.

The framework introduces two key components:

• WT-Aug: A wavelet-based feature augmentation module to enhance feature diversity.
• CG-Fuse: A contextual guidance fusion module that effectively integrates semantic and spatial information.

DINO-AugSeg is extensively evaluated on multiple public medical imaging benchmarks spanning MRI, CT, ultrasound, endoscopy, and dermoscopy, demonstrating robust cross-modality generalization in few-shot scenarios.

---

✨ Features

• Few-shot medical image segmentation with self-supervised DINOv3 features
• Feature-level augmentation using wavelet transforms (WT-Aug)
• Multi-scale contextual fusion via CG-Fuse
• Supports multiple imaging modalities and lesion types
• Modular, readable training and evaluation pipelines

---

🚀 How to Run

1. Clone the Repository

git clone https://github.com/apple1986/DINO-AugSeg.git
cd DINO-AugSeg

2. Download Required Data and Models

Download the processed ACDC dataset and pretrained models from:

🔗 https://drive.google.com/file/d/1QKSINiKGvRRzq2dw2q85HSQ6uZDOmh28/view?usp=sharing

• Unzip the downloaded file.
• Place the model file dinov3_convnext_large_pretrain_lvd.pth in checkpoint/dino_ori
• Place the model file acdc_model_last_epoch.pth in checkpoint/checkpoint_acdc/dino_augseg/cross_guide_wt_unet_7
• Place the ACDC dataset in dataset/ACDC
• Please refer the Organize Files below:

3. Organize Files

DINO-AugSeg/
├── dataset/
│   └── ACDC/
├── checkpoint/
│   ├── dino_ori/
│   └── checkpoint_acdc/
│       └── dino_augseg/
│           └── cross_guide_wt_unet_7/

4. Modify Paths

Update the root paths in the following files to match your local environment:

root_path = "/your/absolute/path"
REPO_DIR = os.path.join(root_path, "DINO-AugSeg")

5. Check the Model

python model/Dinov3_WTAUG_UNet.py

6. Testing

python test_dino_augseg_acdc.py

7. Training

python train_dino_augseg_acdc.py

---

📄 Citation

@article{xu2026dinoaugseg,
  title={Exploiting DINOv3-Based Self-Supervised Features for Robust Few-Shot Medical Image Segmentation},
  author={Xu, Guoping and Udupa, Jayaram K and Lu, Weiguo and Zhang, You},
  year={2026},
  note={arXiv preprint}
}

---

Acknowledgment

This work builds upon recent advances in self-supervised vision models, particularly DINOv3 (https://github.com/facebookresearch/dinov3 ), and the growing body of research on foundation models for medical image analysis. We gratefully acknowledge the developers of the open-source codebases and pretrained models, as well as the contributors of the public datasets used in this study.

📜 License

This project is released under the Apache-2.0 License.