SegDINO: An Efficient Design for Medical and Natural Image Segmentation with DINO-V3

We propose SegDINO, an efficient image segmentation framework that couples a frozen DINOv3 backbone with a lightweight MLP decoder, achieving state-of-the-art performance on both medical and natural image segmentation tasks while maintaining minimal parameter and computational cost.

conda create -n segdino python=3.10.16

conda activate segdino

pip install -r requirements.txt

Clone the DINOv3 repository:

git clone https://github.com/facebookresearch/dinov3.git

Download DINOv3 pretrained weights and place them in:

./web_pth

• TN3K
  A large-scale thyroid nodule segmentation dataset~\citep{gong2023thyroid}, containing 3,493 ultrasound images with pixel-level annotations collected from multiple hospitals.

• Kvasir-SEG
  A polyp segmentation dataset~\citep{jha2019kvasir} derived from colonoscopy examinations, consisting of 1,000 images with high-quality expert annotations.

• ISIC
  A skin lesion segmentation benchmark~\citep{codella2018skin}, providing 2,750 dermoscopic images annotated for lesion boundaries and covering diverse lesion types and acquisition conditions.

Organize datasets in the following structure:

./segdata/tn3k
./segdata/kvasir
./segdata/isic

Each dataset folder should contain an image directory and a mask directory.

Example training command:

python train_segdino.py \
  --data_dir ./segdata \
  --dataset tn3k \
  --input_h 256 --input_w 256 \
  --dino_size s \
  --dino_ckpt ./web_pth/dinov3_vits16_pretrain_lvd1689m-08c60483.pth \
  --repo_dir ./dinov3 \
  --img_dir_name image \
  --label_dir_name mask \
  --mask_ext '.jpg' \
  --epochs 50 \
  --batch_size 4 \
  --lr 1e-4

Example testing command:

python test_segdino.py \
  --data_dir ./segdata \
  --dataset tn3k \
  --input_h 256 --input_w 256 \
  --dino_size s \
  --dino_ckpt ./web_pth/dinov3_vits16_pretrain_lvd1689m-08c60483.pth \
  --ckpt ./tn3k.pth \
  --repo_dir ./dinov3 \
  --img_dir_name image \
  --label_dir_name mask \
  --mask_ext '.jpg'

• Make sure the pretrained DINO weights (.pth file) are correctly downloaded and placed under ./web_pth.
• Modify paths as needed for your environment.
• Training and testing configurations (e.g., image size, batch size, learning rate) can be adjusted via command-line arguments.

We would like to thank the open-source community for their invaluable contributions.
In particular, we acknowledge the following repositories that made this work possible:

• DINOv3

• DPT

• Unimatch

If you find this work useful in your research, please consider citing:

@article{yang2025segdino,
  title={SegDINO: An Efficient Design for Medical and Natural Image Segmentation with DINO-V3},
  author={Yang, Sicheng and Wang, Hongqiu and Xing, Zhaohu and Chen, Sixiang and Zhu, Lei},
  journal={arXiv preprint arXiv:2509.00833},
  year={2025},
  url={https://arxiv.org/abs/2509.00833}
}