---
license: apache-2.0
tags:
- medical
- computer-vision
- image-segmentation
- vascular
- few-shot-learning
size_categories:
- 10K<n<100K
---
# Dataset Card for UniVG-58K

UniVG-58K is a large-scale vascular image dataset containing 58,689 vascular images covering five different medical imaging modalities. This dataset is specifically designed for universal few-shot vascular image segmentation tasks, supporting research and development of generative data-engine foundation models.

## Dataset Details

### Dataset Description

UniVG-58K is a comprehensive vascular image dataset that integrates content from twenty public datasets and one private dataset, ensuring diverse representation of vascular structures. The dataset supports UniVG (Generative Data-engine Foundation Model for Universal Few-shot Planar Vascular Image Segmentation) framework research, achieving efficient vascular segmentation in data-scarce scenarios through compositional learning and few-shot generative adaptation techniques.

The dataset covers five medical imaging modalities:
- **Fundus Imaging**: 42,259 images from seven authoritative data sources
- **Coronary Artery Digital Subtraction Angiography (DSA)**: 12,946 images from three different collections  
- **Brain Digital Subtraction Angiography (DSA)**: 441 images
- **Optical Coherence Tomography Angiography (OCTA)**: 829 high-resolution images
- **Optical Coherence Tomography (OCT)**: 600 images

- **Curated by:** Research teams from Southeast University, Harbin Institute of Technology, Ningbo Institute of Materials Technology and Engineering, Beijing Institute of Technology, French National Institute of Health and Medical Research, Case Western Reserve University, and other institutions
- **Funded by:** National Key Research and Development Program, National Natural Science Foundation of China, Jiangsu Province Natural Science Foundation, and other scientific research funds
- **Shared by:** Paper author team, code  will be publicly available at https://github.com/XinAloha/UniVG
- **Language(s) (NLP):** Not applicable (computer vision dataset)
- **License:** Apache 2.0

### Dataset Sources

- **Repository:** https://github.com/XinAloha/UniVG
- **Paper:** "Generative Data-engine Foundation Model for Universal Few-shot 2D Vascular Image Segmentation", Medical Image Analysis (2025)

## Uses

### Direct Use

The UniVG-58K dataset is designed for the following purposes:

- **Vascular Image Segmentation Research**: Supporting the development of vascular structure segmentation algorithms under various medical imaging modalities
- **Few-shot Learning**: Providing pre-training foundation for vascular segmentation tasks in data-scarce scenarios
- **Generative Model Training**: Supporting vascular image generation and enhancement based on diffusion models
- **Cross-modal Research**: Facilitating knowledge transfer between different medical imaging technologies
- **Foundation Model Development**: Providing large-scale data support for pre-training medical image analysis foundation models
- **Education and Research**: Providing standardized datasets for medical image processing courses and research projects

### Out-of-Scope Use

The following uses are not within the dataset design scope:

- **Direct Clinical Diagnosis**: The dataset is for research purposes only and should not be used directly for clinical diagnosis or treatment decisions
- **Commercial Medical Products**: Should not be used for commercial medical devices or software without proper validation and regulatory approval
- **Non-vascular Structure Segmentation**: The dataset is specifically for vascular structures and is not suitable for segmentation tasks of other anatomical structures

## Dataset Structure

The structure of the UniVG-58K dataset is as follows:

### Data Splits
- **Pre-training Dataset**: 57,075 images for large-scale pre-training of foundation models
- **Downstream Task Dataset**: 1,614 images with manually annotated segmentation masks for few-shot learning evaluation

### Modality Distribution
| Imaging Modality | Pre-training Images | Downstream Task Images | Total |
|------------------|---------------------|------------------------|-------|
| Fundus Imaging | 42,259 | 304 | 42,563 |
| Coronary Artery DSA | 12,946 | 426 | 13,372 |
| Brain DSA | 441 | 284 | 725 |
| OCTA | 829 | 300 | 1,129 |
| OCT | 600 | 300 | 900 |

### Data Format
- **Image Format**: JPEG/PNG, with resolutions ranging from 433×289 to 5184×3456 pixels
- **Annotation Format**: Binary segmentation masks with pixel-level vascular structure annotations
- **File Organization**: Hierarchically organized by modality and task type

### Source Data

#### Data Collection and Processing

The dataset integrates vascular images from multiple authoritative sources:

**Fundus Image Data Sources (7)**:
- DiaRetDB1, e-Ophtha, IDRID, MESSIDOR, Refuge, APTOS 2019, Eyepacs
- Total of 42,259 images covering different imaging conditions and pathological features

**Coronary Artery DSA Data Sources (3)**:
- XACD, ARCADE, SBCD
- Acquired using standard angiographic techniques with resolutions from 512×512 to 1024×1024 pixels

**Brain DSA Data Sources (2)**:
- DIAS, DSCA
- Contains brain angiographic data with different resolutions

**OCT/OCTA Data Sources (2)**:
- OCTA-500, ROSE
- Provides co-registered structural OCT and angiographic OCTA data

All images have been preprocessed to ensure consistency in resolution and format, with quality control and annotation validation by senior medical experts.

#### Who are the source data producers?

The source data of the dataset comes from multiple institutions:

- **Clinical Institutions**: French medical centers, Aravind Eye Hospital in India, Dutch diabetic retinopathy screening programs, University of Calgary clinical facilities in Canada, Shenzhen Second People's Hospital, etc.
- **Research Institutions**: French research institutions, University of Reading in the UK, multiple German medical centers
- **Technical Personnel**: Experienced radiologists, ophthalmologists, and medical image analysis experts
- **Annotation Experts**: Clinicians from collaborating ophthalmology clinics and retinal image analysis experts

All data collection and use follow corresponding ethical approvals and privacy protection protocols.

### Annotations

#### Annotation process

The dataset annotation process includes:

- **Manual Pixel-level Annotation**: Professional physicians perform precise pixel-level segmentation annotations of vascular structures


## Glossary

**Key Term Explanations:**

- **DSA (Digital Subtraction Angiography)**: A medical imaging technique used to visualize vascular structures
- **OCTA (Optical Coherence Tomography Angiography)**: Non-invasive angiographic imaging technique
- **OCT (Optical Coherence Tomography)**: High-resolution medical imaging technique
- **Fundus Imaging**: Imaging technique for examining retinal vascular structures
- **Few-shot Learning**: Machine learning method that trains models using a small amount of annotated data
- **Compositional Learning**: Learning method that generates new samples through decomposition and recombination of entities
- **Diffusion Model**: Generative model used for image generation and editing
- **DSC (Dice Similarity Coefficient)**: Metric for evaluating segmentation accuracy
- **clDice**: Centerline Dice coefficient, specifically for evaluating the quality of elongated structure segmentation

## More Information

For more detailed information, please refer to:

- **Project Homepage**: https://github.com/XinAloha/UniVG
- **Technical Documentation**: Contains detailed implementation details and experimental configurations
- **Usage Examples**: Provides complete code examples and tutorials
- **Benchmark Testing**: Detailed evaluation results on 11 vascular segmentation tasks

The dataset supports PyTorch framework and is compatible with mainstream deep learning platforms.


**Technical Support:**
- GitHub Issues: https://github.com/XinAloha/UniVG/issues
- Project Homepage: https://github.com/XinAloha/UniVG

For dataset usage questions, technical support needs, or academic collaboration inquiries, please contact the project team through the above channels.