Spatially-adaptive Gaussian perturbation for reversible privacy-preserving medical image sharing

Abstract

The rapid integration of digital technologies into modern healthcare has led to an unprecedented exchange of medical imaging data across clinical and remote platforms, raising critical concerns about patient privacy and data exposure. While conventional encryption techniques ensure secure storage and transmission, protected images become fully vulnerable once decrypted. Moreover, many existing protection schemes fail to guarantee faithful reversibility of the original diagnostic content. We propose a key-controlled, entropy-guided Gaussian perturbation framework for reversible privacy preservation. The method performs local entropy analysis to identify sensitive regions, then injects spatially adaptive Gaussian noise using cryptographically secure pseudo-random sequences from a 256-bit key. High-entropy pathological structures receive amplified perturbation while anatomical contexts are preserved. Exact reversibility is achieved through deterministic inversion. Experiments on ChestX-ray14, BraTS 2021, and OCTID demonstrate: imperceptibility (PSNR 41.8 dB, SSIM 0.971), protection against ResNet-50/U-Net (ASR 8.2%, SRR 28.5%), near-exact reconstruction (MAE 0.14), and real-time processing (57 ms/image). Cross-dataset generalization and JPEG compression robustness confirm practical viability. This training-free method enables secure telemedicine and research data sharing without compromising diagnostic integrity.