DwiMark: a multiscale robust deep watermarking framework for diffusion-weighted imaging images

Abstract

To prevent diffusion-weighted imaging (DWI) images from being illegally used or maliciously attacked during the transmission of a remote diagnosis, the copyright of DWI images urgently needs to be protected. In this paper, an end-to-end robust deep watermarking framework is proposed for DWI images, which combines a generative adversarial structure with multiscale features. First, the DWI images are reconstructed by connecting full-scale features to make the fibers and texture highly similar to the original DWI images. Watermarks are also embedded into the multiscale reconstructed features. Then, an optimized BEGAN discriminator is proposed to improve the convergence speed and the visual quality of the reconstructed image. Finally, pyramid filters and multiscale max-pooling are applied to fully learn the watermark distribution features. The experiments show that the proposed framework achieves higher watermark robustness under various common image distortions. Specifically, for rotation and dropout, the robustness can compete with traditional methods. The average PSNR of watermarked DWI images is 58.69\xa0dB, and the diffusion features change indistinguishably, which can effectively protect them without affecting a doctor’s diagnosis.