A zero-watermark scheme for the protection of pediatric orthopedic medical images using dual-source collaborative cross-attention mechanism

Digital medical images of pediatric orthopedic diseases are widely exchanged and transmitted on medical sharing platforms for medical usage. However, it faces serious copyright challenges due to the ease of replication and the limited effectiveness of copyright protection tools. To solve the problems, this paper pro- poses a deep learning zero-watermark scheme based on a dual-source collaborative cross-attention mechanism. First, shallow features of medical and watermark images are extracted via convolutional neural networks. Then, the proposed dual-source collaborative cross-attention fuses host and watermark features, selectively enhancing relevant information while suppressing noise, thus reducing feature homogenization and improving discriminability. Finally, the fused representation is decoded to construct the zero-watermark image. Experimental results have shown that the reconstructed watermark information remains authentic under filtering, noise, rotation, and crop attacks, achieving ideal Peak Signal-to-Noise Ratio (PSNR) and Normalized Correlation (NC) scores. It has been demonstrated that the proposed method offers technical support for the trustworthy sharing and standardized authentication of medical images related to pediatric orthopedic diseases.