Watermarking algorithms that use convolution neural networks have exhibited good robustness in studies of deep learning networks. However, after embedding watermark signals by convolution, the feature fusion efficiency of convolution is relatively low; this can easily lead to distortion in the embedded image. When distortion occurs in medical images, especially in diffusion tensor images (DTIs), the clinical value of the DTI is lost. To address this issue, a robust watermarking algorithm for DTIs implemented by fusing convolution with a Transformer is proposed to ensure the robustness of the watermark and the consistency of sampling distance, which enhances the quality of the reconstructed image of the watermarked DTIs after embedding the watermark signals. In the watermark-embedding network, T1-weighted (T1w) images are used as prior knowledge. The correlation between T1w images and the original DTI is proposed to calculate the most significant features from the T1w images by using the Transformer mechanism. The maximum of the correlation is used as the most significant feature weight to improve the quality of the reconstructed DTI. In the watermark extraction network, the most significant watermark features from the watermarked DTI are adequately learned by the Transformer to robustly extract the watermark signals from the watermark features. Experimental results show that the average peak signal-to-noise ratio of the watermarked DTI reaches 50.47 dB, the diffusion characteristics such as mean diffusivity and fractional anisotropy remain unchanged, and the main axis deflection angle α_AC is close to 1. Our proposed algorithm can effectively protect the copyright of the DTI and barely affects the clinical diagnosis.

使用卷积神经网络的水印算法在深度学习网络的研究中表现出良好的鲁棒性。但通过卷积嵌入水印信号后，卷积的特征融合效率较低；这很容易导致嵌入图像失真。当医学图像，特别是弥散张量图像（DTI）出现失真时，DTI的临床价值就会丧失。针对这一问题，提出了一种通过融合卷积和Transformer实现的DTI鲁棒水印算法，以保证水印的鲁棒性和采样距离的一致性，从而提高了嵌入水印后水印DTI的重建图像的质量信号。在水印嵌入网络中，T1加权（T1w）图像被用作先验知识。提出了 T1w 图像与原始 DTI 之间的相关性，以利用 Transformer 机制从 T1w 图像中计算出最显着的特征。将相关性的最大值作为最显着的特征权重，以提高重建DTI的质量。在水印提取网络中，Transformer充分学习来自加水印的DTI的最重要的水印特征，以从水印特征中鲁棒地提取水印信号。实验结果表明，水印DTI的平均峰值信噪比达到50.47 dB，平均扩散率和分数各向异性等扩散特性保持不变，主轴偏转角α _AC接近1。我们提出的算法可以有效保护DTI的版权，且几乎不影响临床诊断。