Foundation Models in Radiology: What, How, Why, and Why Not

Area of Science:

• Artificial Intelligence
• Deep Learning
• Medical Imaging

Background:

• Large-scale deep learning models, known as foundation models (FMs), excel at interpreting and generating textual and imaging data.
• FMs are trained on vast unlabeled datasets, demonstrating high performance across diverse tasks and attracting significant attention.
• The transformative potential of FMs in radiology necessitates awareness among radiologists regarding their training and application.

Purpose of the Study:

• To provide radiologists with a foundational understanding of FMs in the context of radiology.
• To elucidate the requirements for training radiology-specific FMs, including data, paradigms, capabilities, and evaluation.
• To bridge the gap between technical advancements in FMs and clinical needs for safe and responsible implementation.

Main Methods:

• Review of fundamental concepts and terminology related to FMs in radiology.
• Focused discussion on essential components for training radiology-specific FMs: data requirements, training paradigms, model capabilities, and evaluation strategies.
• Synthesis of technical advancements with clinical requirements for responsible FM development.

Main Results:

• Detailed explanation of foundation model concepts tailored for a radiology audience.
• Identification of key considerations for training radiology-specific FMs, encompassing data curation, algorithmic approaches, performance assessment, and ethical implications.
• Framework for aligning AI development with clinical practice to ensure patient benefit.

Conclusions:

• Radiologists need to understand foundation models to leverage their potential in medical imaging.
• A comprehensive approach to training radiology-specific FMs is crucial, balancing technical capabilities with clinical utility and safety.
• Responsible development and deployment of FMs in radiology promise to enhance patient care and provider efficiency.