Privacy-Preserving Medical Image Steganography: A Clinical Quality-Aware Deep Learning Framework

We propose a deep learning framework for privacy-preserving medical image steganography using a convolutional neural network (CNN)-based architecture. Our framework consists of an encoder that generates minimal residual perturbations for embedding secret medical images, and a decoder for extraction. The encoder employs parallel processing paths with 32-filter convolutional layers for feature extraction from both cover and secret images, followed by channel-wise concatenation and residual generation scaled by 0.1. The decoder implements a progressive feature expansion-compression strategy with filter dimensions [32, 64, 128, 64, 1]. We evaluate the framework on the MIDRC-RICORD-1B and IQ-OTH/NCCD datasets, achieving average PSNR values of 44.31 dB for cover-stego pairs and 37.57 dB for secret-recovered pairs, with corresponding SSIM values of 0.9626 and 0.9897. Empirical analysis demonstrates the framework’s ability to preserve diagnostic quality while maintaining embedding imperceptibility, significantly outperforming existing methods by 9.76 dB in stego image quality. Our architecture supports full-sized medical images at 128×128 resolution with a 100% embedding capacity.