基于注意力机制U-Net的低剂量CT图像去噪方法

Low-dose CT Image Denoising Basedon U-Net with Attention

计算机断层扫描(CT)产生的辐射风险已成为公众关注的问题。降低剂量将影响CT图像的质量以及医生的诊断结果。传统的基于深度网络算法中,同一层中的特征通道间的地位是平等的,影响信息的提取。为此,提出了一种具有注意力机制的U-Net残差网络。在U-Net中引入通道注意力模块驱使网络将更多的注意力集中于含有噪声和伪影信息的通道上,降低对其他信息通道的关注度,提高网络的去噪能力。同时利用像素注意力引导网络捕捉图像区域中的噪声和伪影。实验验证了两种注意力的结合提高了网络的性能。所提出的方法的峰值信噪比高出REd-CNN大约0.17,视觉信息保真度高出大约5%,信息保真度准则高出大约4%.

The radiation risk from computed tomography(CT) has become a public concern.Reducing the dose will influence the quality of CT images and the doctors’ diagnosis.In the methods based on traditional deep network,the status of feature channels in the same layer is equal,which affects the extraction of information.In order to solve the problem,a U-Net residual network with attention mechanism was proposed.The channel attention module was introduced to drive the network to focus more attention on the channels containing noise and artifact information,reduce the attention to other information channels,and improve the denoising ability of the network.At the same time,the pixel attention was used to guide the network to capture noise and artifacts in the image.Experiments verified that the combination of two attention modules improved the network performance.The peak signal-to-noise ratio of the proposed method is about 0.17 higher than that of REd-CNN,the visual information fidelity is about 5% higher,and the information fidelity criterion is about 4% higher.