Cautious optimism on foundation models in medical imaging balancing privacy and innovation

Key Points

• This research examines the potential and risks of foundation models in medical imaging, particularly relating to privacy concerns.
• Analysis of re-identification rates in retinal imaging using foundation models.
• Evaluation of the risks associated with latent demographic signals during model training.
• Discussion of technical safeguards and policy frameworks to enhance privacy.
• Re-identification rates reached up to 94 percent in retinal imaging with foundation models.
• Identified the need for strategies to mitigate privacy risks and protect patient confidentiality.