Foundation models in biomedical imaging: turning hype into reality

Area of Science:

• Biomedical Imaging AI
• Foundation Models (FMs)
• Clinical Translation

Background:

• Foundation models (FMs) offer unified backbones for diverse biomedical imaging tasks, enabling data integration.
• This contrasts with medicine's sub-specialization, creating tension with FM integration.
• Data scarcity, domain heterogeneity, and limited interpretability hinder real-world clinical value.

Purpose of the Study:

• To evaluate the real-world readiness of FMs in biomedical imaging.
• To introduce a framework (REAL-FM) for assessing FMs across multiple dimensions.
• To differentiate hype from clinical utility of FMs.

Main Methods:

• Developed and applied the REAL-FM framework.
• Assessed FMs on data, technical readiness, clinical value, workflow integration, and responsible AI.
• Analyzed FM performance in pattern recognition, causal reasoning, domain robustness, and safety.

Main Results:

• FMs excel at pattern recognition but lack causal reasoning, domain robustness, and safety.
• Clinical translation is limited by scarce data, unverified generalization, and lack of outcome-based validation.
• Human oversight remains indispensable for safe and effective clinical use.

Conclusions:

• FMs should augment, not replace, clinical expertise in biomedical imaging.
• A multi-dimensional framework (REAL-FM) is crucial for evaluating FM clinical readiness.
• Future AI in medicine requires coordinated, transparent, safe, and clinically grounded subspecialist systems.