【摘要】 为了提高去噪网络的可解释性，将传统滤波算子的优势融入到网络设计中，提出了基于跨尺度边缘增强的深度卷积神经网络(CEDCNN).将传统边缘算子与卷积相结合，设计出轻量化的边缘增强模块，增强边缘信息对网络结果的影响.基于自适应一致性先验算法构建深度迭代网络，进一步提取边缘增强特征，从而实现端到端可训练、可解释的深度去噪网络.将均方误差和多尺度注意残差感知损失相结合，解决了重建图像过平滑的问题.实验结果表明，CEDCNN去噪网络的PSNR、RMSE、SSIM评价指标分别为43.647 5 dB、0.006 8、0.987 5,说明该方法可显著提高去噪后的图像质量，有效保证低剂量X射线下CT图像的成像质量和精度，去噪效果与正常剂量CT图像所展现的人体组织细节相当.更多还原

【Abstract】 To improve the interpretability of denoising network, the advantages of traditional filtering operators were incorporated into network design, and a cross-scale edge-enhanced deep convolutional neural network(CEDCNN) was proposed. The traditional edge operator was combined with convolution to design a lightweight edge enhancement module, which enhanced the influence of edge information on the network results. A deep iterative network was constructed based on an adaptive consistency prior algorithm, and edge enhancement features were further extracted to achieve an end-to-end trainable and interpretable deep denoising network. The composite loss combining mean square error(MSE) and multi-scale attention residual perception were combined to solve the problem that the reconstructed images were easily over-smoothed. The experimental results show that the peak signal to noise ratio(PSNR), the root mean square error(RMSE), the structural similarity(SSIM) of the CEDCNN denoising network are 43.647 5 dB, 0.006 8, 0.987 5, respectively. The image quality after denoising is significantly improved. The imaging quality and accuracy of the CT images can be effectively guaranteed at low dose of X-ray. The denoising results are comparable to the details of human tissues shown in normal dose CT image.更多还原