A foundation generative model for breast ultrasound image analysis

Area of Science:

• Artificial Intelligence in Medical Imaging
• Foundation Models for Healthcare
• Breast Ultrasound Analysis

Background:

• Foundation models offer significant potential for clinical applications, but their use in breast ultrasound image analysis is underexplored.
• Developing robust AI models for breast ultrasound requires large, diverse datasets, which are often difficult to obtain.
• Existing methods lack specialized foundation models for the nuances of breast ultrasound data.

Purpose of the Study:

• Introduce BUSGen, the first foundation generative model specifically designed for breast ultrasound image analysis.
• Leverage a large dataset of breast ultrasound images to train a versatile foundation model.
• Demonstrate BUSGen's capability to generate high-quality synthetic data for downstream AI tasks.

Main Methods:

• Pretrained BUSGen on over 3.5 million breast ultrasound images.
• Utilized few-shot adaptation to generate task-specific synthetic data.
• Conducted extensive experiments comparing BUSGen-generated data with real data for various breast cancer analysis tasks.

Main Results:

• BUSGen demonstrated exceptional adaptability, outperforming real-data-trained foundation models in screening, diagnosis, and prognosis.
• In early breast cancer diagnosis, BUSGen achieved a 16.5% average sensitivity improvement, surpassing nine board-certified radiologists.
• Characterized the scaling effects of synthetic data and enabled secure, de-identified data sharing.

Conclusions:

• BUSGen represents a significant advancement in AI for breast ultrasound, enabling improved diagnostic accuracy and efficiency.
• The model's ability to generate realistic synthetic data addresses data scarcity and privacy concerns in medical AI.
• BUSGen facilitates secure medical data utilization and accelerates the development of AI tools for breast cancer detection and management.