Generative AI enables medical image segmentation in ultra low-data regimes.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning excels at medical image segmentation for diagnosis and treatment planning.
• However, deep learning models require large annotated datasets, which are scarce in medical imaging.
• This data scarcity limits the application of deep learning in ultra low-data regimes.

Purpose of the Study:

• To develop a generative deep learning framework for creating high-quality medical image-mask pairs.
• To address the challenge of limited annotated data in medical image segmentation.
• To improve the feasibility and cost-effectiveness of deep learning in data-limited medical imaging.

Main Methods:

• A novel generative deep learning framework utilizing multi-level optimization for end-to-end data generation.
• The framework integrates data generation directly with the segmentation model training process.
• Segmentation performance guides the data generation, creating tailored auxiliary training data.

Main Results:

• The generative framework demonstrated strong generalization across 11 medical image segmentation tasks and 19 diverse datasets.
• Significant performance improvements of 10-20% (absolute) were observed in both same- and out-of-domain settings.
• The method requires 8-20 times less training data compared to existing approaches.

Conclusions:

• The proposed generative framework effectively overcomes data scarcity in medical image segmentation.
• It enhances deep learning model performance and generalization in low-data scenarios.
• This approach significantly improves the practicality and cost-efficiency of AI in medical imaging.