A Survey on Trustworthiness in Foundation Models for Medical Image Analysis

Foundation models in medical imaging need rigorous trustworthiness evaluation across five critical dimensions: privacy, robustness, reliability, explainability, and fairness—yet existing research has significant gaps in addressing these concerns, particularly in how they apply specifically to medical imaging applications.

Four major medical imaging applications are leading adoption of foundation models: segmentation (identifying anatomical structures), medical report generation, medical question-and-answering systems, and disease diagnosis—but each presents distinct trustworthiness challenges that require tailored solutions rather than one-size-fits-all approaches.

Current surveys on foundation models fail to adequately address trustworthiness concerns specific to healthcare, creating a dangerous gap between rapid deployment of these powerful AI systems and our understanding of whether they can be safely and fairly used in clinical practice.

The stakes of deploying untrustworthy foundation models in medical imaging are exceptionally high: diagnostic errors, privacy breaches, or biased recommendations could directly harm patients, making trustworthiness not just a technical concern but an ethical imperative for healthcare AI.

Building trustworthy medical imaging AI requires examining complex, interconnected challenges: the survey identifies that trustworthiness concerns vary significantly across different medical imaging tasks, suggesting that healthcare institutions need application-specific strategies rather than generic AI governance frameworks.