ROFI: a deep learning-based ophthalmic sign-preserving and reversible patient face anonymizer

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Patient face images are crucial for diagnosing eye diseases but raise significant privacy concerns.
• Existing methods may compromise diagnostic accuracy when anonymizing sensitive facial data.

Purpose of the Study:

• To introduce ROFI, a novel deep learning framework for privacy protection in ophthalmology image analysis.
• To evaluate ROFI's ability to anonymize facial features while retaining critical disease indicators.

Main Methods:

• ROFI utilizes weakly supervised learning and neural identity translation for facial anonymization.
• The framework was tested for diagnostic accuracy, diagnostic sensitivity, and image anonymization rates across multiple cohorts and eye diseases.

Main Results:

• ROFI achieved over 98% accuracy and κ > 0.90 in retaining disease features.
• It demonstrated 100% diagnostic sensitivity and high agreement (κ > 0.90) for eleven eye diseases, anonymizing over 95% of images.
• The framework maintained original diagnoses (κ > 0.80) when integrated with AI systems and allowed secure image reversal.

Conclusions:

• ROFI effectively anonymizes patient images in ophthalmology, safeguarding privacy without sacrificing diagnostic integrity.
• The framework supports AI integration and secure data handling, crucial for modern digital healthcare.
• ROFI represents a significant advancement in protecting patient privacy within the digital medicine era.