A Robust and Blind Watermarking Framework for Medical Images Using Whale-Optimized DCT-QIM with CNN Restoration

This paper presents a robust and imperceptible watermarking scheme for DICOM medical images that integrates the Whale Optimization Algorithm (WOA) with a Denoising Convolutional Neural Network (CNN). The objective is to embed sensitive patient information directly inside the medical image so that authenticity, confidentiality, and integrity are preserved without degrading diagnostic quality. In the proposed scheme, each DICOM slice is segmented to isolate its Region of Interest (ROI), which is then partitioned into equal-sized blocks and transformed into the frequency domain using the block Discrete Cosine Transform (DCT). A chaotic Duffing map generates a secret key that determines the candidate blocks used for embedding, while the WOA optimizes the per-slice quantization step of a Quantization-Index-Modulation (QIM) embedding to minimize the average Mean Squared Error (MSE) while preserving robustness, thereby achieving the best trade-off between imperceptibility and robustness. After embedding, a Denoising CNN suppresses the residual distortion introduced during the embedding stage, and the processed slices are merged back into the watermarked DICOM volume. The watermark is blindly extracted by minimum-distance QIM decoding of the same DCT coefficients followed by majority voting over the redundant copies, requiring no access to the original image. Experiments on a real 10-slice brain MRI volume (256×256 per slice) show that the method achieves an average Peak Signal-to-Noise Ratio (PSNR) of 55.66 dB and a Structural Similarity Index (SSIM) of 0.9993, with lossless recovery of the embedded patient record on the clean image (bit-error rate, BER = 0). The scheme further withstands JPEG compression (quality ≥ 70), mild Gaussian noise, and corner cropping, while a Denoising CNN restores up to 2 dB of visual quality on noise-corrupted received images.