Research On Robust Watermarking Algorithm For Diffusion Magnetic Resonance Tensor Imagin

Digital watermarking of medical images has become an effective means of protecting image integrity,limiting unauthorized use and providing accurate images to remote experts.Watermarking methods based on deep learning have been successfully applied to copyright protection of real-value domain medical images such as Computed Tomography(CT),Magnetic Resonance Imaging(MRI)and Diffusion Weighted Imaging(DWI).However,manifold-value domain images,such as diffusion tensor imaging(DTI),have received less academic attention.DTI is a noninvasive inspection method that can play an important role in clinical applications,but its non-Euclidean nature hinders deep neural network-based watermarking algorithms.In this paper,we propose two deep learning-based robust watermarking algorithms for DTI images,aiming to solve this problem,which are as follows:1.Robust watermarking algorithm for DTI images based on voxel space transformation.First,a voxel space transformation is proposed for the first time,using singular value decomposition to transform the voxels of DTI from the 3×3 symmetric positive definite(SPD)manifold space to Euclidean space and obtain the eigenvalues and eigenvectors of each voxel.Second,SVD is combined with a deep neural network to extract the high-level features of each voxel eigenvalue in the DTI images and embedding watermarking messages.Finally,using a deep neural network combining multiscale dilated convolution,dense residual connection and channel attention,the algorithm demonstrates high robustness against intentional or unintentional attacks on DTI images while ensuring the good visual quality and diffusion characteristics of the DTI images embedded with watermarking messages.2.Two-stage Separable Adversarial Distortion robust watermarking framework for DTI images.The first phase uses a noise-free end-to-end watermark embedding and extraction network for learning and training of high-dimensional features.In the second stage,the watermark embedding network trained in the first stage is fixed,and the noisy distortion network and the watermark extraction network are interacted to perform the adversarial training.The most aggressive distortion simulated by the adversarial distortion network is used to perform noise attacks on the watermarked DTI images by replacing the known distortion with the most aggressive distortion to generate the corresponding adversarial training set,and then the adversarial training is used to improve the robustness of the watermark extraction network in the face of unknown distortion.