Generative Data-engine Foundation Model for Universal Few-shot 2D Vascular Image Segmentation

A research team led by Rongjun Ge and 10 collaborators has introduced UniVG (Generative Data-engine Foundation Model for Universal Few-shot 2D Vascular Image Segmentation), a breakthrough AI model that solves one of medical imaging's biggest challenges: the scarcity of annotated vascular data. Traditional deep learning approaches require thousands of labeled images to segment blood vessels accurately, but UniVG achieves comparable performance using just 5 annotated samples per task through two key innovations. First, it uses compositional learning to decompose and recombine vascular structures with varying morphological features, generating diverse synthetic image-label pairs. Second, it employs few-shot generative adaptation to bridge the gap between synthetic and real vascular domains, fine-tuning pre-trained models with minimal real data.

The researchers created UniVG-58K, a massive dataset of 58,689 vascular images across five imaging modalities (including ultrasound, MRI, and CT), to enable robust large-scale generative pre-training. In extensive experiments across 11 vessel segmentation tasks, UniVG demonstrated performance matching fully supervised models while requiring 99.9% less annotated data. This represents a paradigm shift for medical AI deployment, where data annotation costs and privacy concerns have historically limited real-world application. The model's ability to work across multiple imaging modalities makes it particularly valuable for clinical settings where equipment varies.

All code and datasets will be made publicly available, potentially accelerating vascular research worldwide. The approach could extend beyond vascular imaging to other medical segmentation tasks facing similar data scarcity challenges, fundamentally changing how AI models are trained for specialized medical applications.