Towards generalizable AI in medicine via Generalist-Specialist Collaboration

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Healthcare
• Clinical Decision Support Systems

Background:

• Generalist foundation models (GFMs) offer broad capabilities but may lack medical specificity.
• Specialist AI models provide domain-specific expertise but are limited in scope.
• Integrating generalist and specialist models is crucial for advancing medical AI.

Purpose of the Study:

• To introduce the Generalist-Specialist Collaboration (GSCo) framework for medical AI.
• To develop MedDr, a generalist foundation model for medicine.
• To evaluate the performance of GSCo in diverse medical tasks.

Main Methods:

• Developed a cooperative framework (GSCo) combining a generalist model (MedDr) with lightweight specialists.
• Specialist models provide diagnostic predictions and similar case information to guide the generalist.
• Evaluated MedDr and GSCo on 32 datasets across various medical modalities.

Main Results:

• MedDr outperformed existing state-of-the-art GFMs on downstream medical tasks.
• The GSCo framework surpassed both GFMs and specialist models in medical image diagnosis and report generation.
• Demonstrated improved accuracy and efficiency in clinical AI applications.

Conclusions:

• The GSCo framework offers an effective and computationally efficient paradigm for deploying GFMs in clinical settings.
• This approach enhances scalability and enables precise analysis across diverse medical scenarios.
• Cooperative generalist-specialist models represent a promising direction for the future of medical AI.