Anonymizing medical case-based explanations through disentanglement.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning models offer valuable insights into clinical decision-making through case-based explanations.
• Sharing medical images for these explanations raises significant patient privacy concerns.
• Existing anonymization methods may compromise the integrity of medical information.

Purpose of the Study:

• To develop a novel method for disentangling identity and medical characteristics in images.
• To anonymize medical images while preserving their diagnostic content for privacy-preserving explanations.
• To enable the generation of realistic synthetic identities for anonymized medical data.

Main Methods:

• A disentanglement mechanism was designed to separate identity and medical feature vectors within images.
• A model was developed to generate synthetic, privacy-preserving identities to replace original image identities.
• The proposed models were applied to both medical and biometric datasets for validation.

Main Results:

• The method successfully anonymized medical images, generating realistic outputs that retained original medical information.
• The disentanglement process effectively separated identity from medical characteristics.
• The models demonstrated the ability to generate counterfactual images by altering medical features.

Conclusions:

• The proposed approach effectively anonymizes medical images, addressing privacy concerns in AI-driven clinical decision support.
• This technique facilitates the use of case-based explanations with sensitive medical data.
• The method holds potential for generating synthetic medical data and exploring counterfactual scenarios.