Generative AI for developing foundation models in radiology and imaging: engineering perspectives.

Area of Science:

• Artificial intelligence
• Medical imaging
• Foundation models

Background:

• Annotated data in radiology is limited and heterogeneous.
• Generative AI offers solutions for self-supervised learning and synthetic data generation.
• Generative AI addresses scalability, multimodal alignment, and data diversity challenges in medical AI.

Purpose of the Study:

• To review the role of generative AI in medical foundation models.
• To explore generative model frameworks and representation learning techniques.
• To describe multimodal large language models (MLLMs) for clinical applications.

Main Methods:

• Review of generative models (VAEs, diffusion, autoregressive frameworks).
• Exploration of hybrid designs and representation learning (masked autoencoding, contrastive learning).
• Description of MLLM design and training for integrating visual, textual, and clinical data.

Main Results:

• Generative AI models form the backbone of medical foundation models.
• Hybrid designs and representation learning enhance model performance.
• MLLMs integrate diverse data for applications like report generation and clinical reasoning.

Conclusions:

• Generative AI enables scalable, adaptable, and privacy-conscious medical AI.
• Case studies demonstrate the practical application of these models.
• Future directions include addressing hallucination, generalization, and regulatory challenges for trustworthy AI deployment.