基于全卷积金字塔残差网络的能谱CT图像降噪研究

传统CT采用积分式探测器采集投影数据，反映的是物体的平均衰减特性，会在一定程度上造成信息损失，无法对物体进行较好的定性定量测量。基于光子计数探测器的能谱CT通过设定多个能量响应阈值能够探测不同能量范围内的X射线光子，采集更多被测物体的物质组成信息，有助于识别不同物理特性的材料，基于此，能谱CT被广泛的应用于小病灶、低对比度结构以及微细结构的成像。然而将整个能谱划分为多个能量段进行数据采集时，范围较窄能量范围内的有效光子数比例相对降低，导致图像中包含较多的噪声，图像质量较差，影响能谱CT的临床应用。为了有效的抑制能谱CT不同能量段内重建图像中的噪声，提出了一种基于深度学习的能谱CT图像降噪方法。我们将全卷积网络和金字塔残差网络结合为全卷积金字塔残差网络(FCPRN)，实验中，利用能谱CT在不同的能量范围扫描小鼠样本，使用FDK算法和基于压缩感知的Split-Bregman算法进行重建并分别作为训练数据和标签数据训练全卷积金字塔残差网络。为了验证网络的降噪性能，选取了常见的降噪网络模型denoising convolutional neural networks(DNCN)以及residual encoder decoder convolutional neural network (REDCNN)进行对比，训练三种网络的使用的数据和实验配置都是完全相同的，实验结果表明训练模型可以有效抑制不同能量范围内重建图像的噪声，且使用的全卷积金字塔残差网络的降噪性能优于其他网络模型。模型训练好后，可以对FDK算法重建出的图像进行降噪，由此提高能谱CT图像降噪效率，保证能谱CT重建图像的质量。

Research on Spectral CT Image Denoising Via Fully Convolution Pyramid Residual Network

Traditional computed tomography(CT) uses an integral detector to collect projection, which reflects the average attenuation characteristics of the object and causes the loss of attenuation characteristics to some extent, so it cannot measure the object qualitatively and quantitatively. The spectral CT based on photon-counting detectors can collect the incident photons in different energy ranges by setting several energy thresholds to collect more material composition information of measured objects, which is helpful to identify materials with different physical characteristics, so the spectral CT is widely used in imaging of small lesions, low contrast structures and fine structures. However, dividing the whole energy spectrum into several energy segments for data acquisition will lead to the relatively reduced proportion of effective photons, resulting in more noise in the image and affecting the clinical application of energy spectrum CT. To effectively suppress the noise in different energy segments of spectral CT image, we propose an image denoising method basedondeeplearning. We combine the full convolution network and the residual pyramid network into the full convolution pyramid residual network (FCPRN). Our study, scanned a mouse specimen with spectral CT based on photon-counting detector and used the FDK algorithm and Split-Bregman algorithm for reconstruction to obtain training data and labeled data, respectively. Then we use the data set to train our network for image denoising. To verify our network’s performance, we selected the common denoising networks, denoising convolutional neural networks（DNCNN）and residual encoder-decoder convolutional neural network（REDCNN）for comparison, and the training data and experimental configuration of the three networks are identical. Experimental results demonstrated that the proposed method could reduce the noise of spectral CT images in different energy ranges，and the performance of FCPRN is better than that of other neural networks discussed in this paper for denoising. When the model is trained, the image reconstructed by the FDK algorithm can be processed quickly via the model to improve the denoising efficiency and ensure the reconstructed image’s quality of spectral CT.