PRIMIS: Privacy-preserving medical image sharing via deep sparsifying transform learning with obfuscation

Area of Science:

• Medical imaging
• Data privacy
• Machine learning

Background:

• Sharing medical images is crucial for clinical and research purposes.
• Privacy concerns and data security restrictions hinder effective medical image sharing.
• Existing methods often fail to balance data utility with robust privacy preservation.

Purpose of the Study:

• To design a privacy-preserving mechanism for sharing medical images securely.
• To enable storage of obfuscated medical images in the public domain.
• To allow authorized users to reconstruct original images using private keys.

Main Methods:

• Utilized a neural auto-encoder architecture.
• Employed sparsifying transformations to generate multiple compact image codes.
• Implemented obfuscation via pseudo-random noise for privacy preservation.

Main Results:

• Demonstrated effectiveness using chest X-ray images for classification, segmentation, and texture analysis.
• Evaluated the framework's robustness against supervised and unsupervised attacks.
• Confirmed the computational infeasibility for attackers to reconstruct private image data.

Conclusions:

• Introduced a novel, optimized, and privacy-assured medical image data-sharing mechanism.
• Enabled secure multi-party sharing of sensitive medical imaging data.
• The proposed method is adaptable to various medical imaging modalities beyond X-rays.