A Hybrid Deep Learning and Fusion Domain Framework for Secure, High-Capacity Reversible Data Hiding in Color Medical Images

The increasing reliance on digital networks for transmitting medical images in the smart healthcare sector introduces critical challenges related to patient confidentiality, data integrity, and content authenticity. While reversible data hiding (RDH) offers a compelling solution by allowing for the recovery of the original image after secret data extraction, existing approaches often struggle to simultaneously achieve high embedding capacity, robust security, and excellent visual fidelity—a triad of requirements essential for clinical applications. This paper proposes RDHNet, a novel framework designed to address these limitations through a synergistic integration of deep learning, advanced transform techniques, and cryptographic encryption.Our method leverages a pre-trained AlexNet model to extract a robust and semantically rich feature vector from the host color medical image. This feature vector is subsequently transformed into a topographic map using the watershed transform (WST), a process that refines the embedding space. To ensure a high level of security, the transformed features are encrypted using an L-shaped fractal Tromino cryptosystem. The final embedding of the secret data is performed in the transformed domain using a histogram-based shifting strategy, which is pivotal for maintaining high payload capacity while minimizing distortion . The efficacy of RDHNet is rigorously evaluated through extensive experimentation. The results demonstrate that the proposed method achieves an outstanding visual quality, with an average Peak Signal-to-Noise Ratio (PSNR) of 73.14 dB and a Structural Similarity Index Measure (SSIM) of 0.9999, indicating near-perfect reversibility . Furthermore, the framework exhibits exceptional robustness against a wide range of geometric distortions, noise-adding attacks, and common steganalysis techniques, as evidenced by perfect Normalized Correlation (NC = 1) and zero Bit Error Rate (BER = 0) values under normal conditions . By harmonizing deep learning for robust feature extraction, watershed transform for domain manipulation, fractal encryption for security, and histogram shifting for capacity control, RDHNet presents a comprehensive and effective solution for securing sensitive medical imagery in modern healthcare information systems.