Blind and reversible RONI watermarking of medical images using a perceptual CNN and chaotic block shuffling

The transmission of medical images over public networks exposes patient data to unauthorized access and undetected alteration. Existing watermarking techniques either require manual selection of diagnostically relevant regions, rely on non-blind extraction, use lossy compression that corrupts embedded information after tampering, or lack reversibility. This paper introduces a blind and reversible watermarking framework that embeds a QR code exclusively into the non-diagnostic region of medical images. The non-diagnostic region (RONI) is automatically identified using an active contour model without edge constraints, which separates anatomical structures from background regions based on intensity homogeneity. A lightweight convolutional neural network predicts a block-wise adaptive embedding strength in the discrete wavelet transform-singular value decomposition domain, and a chaotic permutation based on the logistic map shuffles the blocks before insertion. After extraction, the original non-diagnostic region is perfectly recovered by reversing the singular value modifications using the same CNN-predicted strengths, without requiring a location map or side information. The key contributions include automatic ROI segmentation eliminating manual intervention, a perceptually adaptive CNN-based embedding strength that improves robustness, and a chaotic block shuffling mechanism that resists cropping and desynchronization attacks. Experimental evaluation on CT, MRI, and X-ray images demonstrates that the proposed method achieves an average peak signal-to-noise ratio of 49.3 dB, a structural similarity index of 0.992, and normalized correlation values above 0.92 under JPEG compression, Gaussian noise, cropping, rotation, and combined attacks, with an embedding time of 0.80 s on a standard CPU. The QR decoding success rate reaches 100 % under moderate attacks and remains above 90 % under severe distortions, confirming the practical utility of the approach for clinical telemedicine applications.