Rethinking Privacy in Medical Imaging AI: From Metadata and Pixel-Level Identification Risks to Federated Learning...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Data Privacy

Background:

• Traditionally, de-identification focused on metadata in medical imaging datasets.
• Deep learning models can extract sensitive patient information from pixel-level image data.
• This poses significant privacy risks beyond traditional metadata concerns.

Purpose of the Study:

• To discuss metadata and pixel-level data as sources of identifiable information in medical imaging.
• To review privacy-preserving techniques like federated learning and synthetic data generation.
• To highlight the limitations and vulnerabilities of these methods in AI applications.

Main Methods:

• Review of existing literature on medical imaging de-identification.
• Analysis of privacy risks associated with metadata and intrinsic image data.
• Evaluation of federated learning and synthetic data generation for privacy preservation.
• Discussion of vulnerabilities such as model inversion and inference attacks.

Main Results:

• Both metadata and pixel-level image data contain identifiable information.
• Federated learning and synthetic data generation offer privacy benefits but are not foolproof.
• Model inversion and inference attacks remain significant threats to privacy.

Conclusions:

• Comprehensive de-identification strategies must address both metadata and image content.
• Limitations of current privacy-preserving methods necessitate careful consideration in AI deployment.
• Ongoing research is crucial to mitigate advanced privacy risks in medical imaging AI.