动态可控残差卷积神经网络的低剂量 CT 图像处理

针对计算机断层扫描(Computed Tomography, CT)中因采用低剂量扫描方式,导致图像噪声伪影干扰,尤其 是不同部位噪声和伪影强度存在较大差异这一问题,提出了一种基于动态可控残差的卷积神经网络(DC-ResNet) 算法。 DC-ResNet 的主要思想是在常规残差网络连接中添加一个图像质量指导的控制变量,以允许获取残差的加 权和,从而实现残差特征的动态可控。 所设计的 DC-ResNet 网络是一种包括两个子网络的组合型结构,一个是作 为主干网络的基础子网络,在该子网络中使用全局动态残差块和局部动态残差块来实现低剂量 CT 图像质量的提 高;另一个是作为辅助网络的条件子网络,用来生成基础子网络中不同动态可控残差块的权值,辅助基础子网络的 学习。 通过 Mayo 与 UIH 数据实验验证,其视觉结果表明:处理后的不同部位 CT 图像噪声伪影均能够得到较好的 抑制,并能有效地保留结构细节及组织纹理;量化结果表明:处理后的 CT 图像峰值信噪比( Peak-Signal to Noise Ratio, PSNR)和结构相似性(Structure Similarity, SSIM)均优于对比方法。

Low Dose CT Image Processing Based on Dynamic Controllable Residual Convolution Neural Network

Aiming at the problem of image noise and artifact interference caused by the low-dose scanning method in Computed Tomography ( CT ), especially the large difference in noise and artifact intensity in different parts, a convolutional neural network algorithm based on dynamic controllable residual (DC-ResNet) was proposed. The main idea of DC-ResNet is to add an image quality-guided control variable to the regular residual network connection to allow a weighted sum of residuals to be obtained, thereby achieving dynamic controllability of residual features. The designed DCResNet network was a combined structure including two sub-networks: one was the basic sub-network as the backbone network, in which the global dynamic residual block and the local dynamic residual block were used to achieve the improvement of low-dose CT image quality; the other was a conditional sub-network, as an auxiliary network, which was used to generate the weights of different dynamically controllable residual blocks in the base subnet and assist the learning of the base subnet. Through the experimental verification of Mayo and UIH data, the visual results showed that the noise and artifacts of the processed CT images in different parts were well suppressed, and the structural details and tissue texture were effectively preserved. The quantification results showed that the peak-signal to noise ratio ( PSNR) and structural similarity (SSIM) of the processed CT images were better than those of the comparison methods.