Generative data-engine foundation model for universal few-shot 2D vascular image segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning for 2D vascular segmentation is clinically valuable but limited by scarce annotated data.
• Developing universal few-shot vascular segmentation models is challenging due to extensive training needs and imaging complexities.

Purpose of the Study:

• To introduce UniVG, a generative foundation model for universal few-shot 2D vascular image segmentation.
• To address the data scarcity issue by enabling synthesis of diverse and realistic vascular images.

Main Methods:

• UniVG utilizes compositional learning to synthesize diverse vascular images and labels by recombining structural features.
• It employs few-shot generative adaptation to fine-tune models with minimal data, bridging synthetic and real data domains.
• A large dataset, UniVG-58K (58,689 images across 5 modalities), was created for generative pre-training.

Main Results:

• UniVG achieved performance comparable to fully supervised models on 11 vessel segmentation tasks using only 5 labeled images per task.
• Demonstrated significant reduction in data collection and annotation costs.
• Validated effectiveness across 5 imaging modalities.

Conclusions:

• UniVG offers a robust solution for few-shot vascular segmentation, overcoming data limitations.
• The generative foundation model approach facilitates transferable segmentation across different modalities.
• Publicly available code and datasets encourage further research and application.