基于卷积神经网络生成虚拟平扫CT图像

目的 针对增强CT图像，采用U-Net卷积神经网络（CNN）方法生成虚拟平扫CT图像，比较其与增强CT和实际平扫CT图像的差异。方法 纳入50例于同次检查中接受平扫及增强CT扫描患者，记录其容积CT剂量指数（CTDIvol）。以40例CT数据为训练集，输入增强CT图像后，以对应的平扫CT图像作为输出，用于训练U-Net神经网络；以其余10例数据作为测试集，通过训练完成的U-Net生成虚拟平扫CT图像。比较虚拟平扫CT与实际平扫及增强CT的图像及参数差异。结果 50例平扫CT平均CTDIvol为（11.67±0.51） mGy，增强CT平均CTDIvol为（13.46±0.76） mGy；采用虚拟平扫CT图像可使平均辐射剂量减少46.44%。增强与平扫CT图像的平均绝对偏差（MAE）为（32.28±2.64） HU，结构相似度（SSIM）为0.82±0.05；虚拟平扫CT与平扫CT图像的MAE为（6.72±1.31） HU，SSIM为0.98±0.02。虚拟平扫CT与平扫CT图像所示主要组织的CT值差异均无统计学意义（P均>0.05）。结论 针对增强CT基于U-Net神经网络生成的虚拟平扫CT图像与实际平扫CT图像的一致性较好，可由此减少CT扫描次数、降低辐射剂量。

Generation of virtual plain CT images based on convolutional neural network

Objective To generate virtual plain CT images from enhanced CT using U-Net convolution neural network (CNN), and to compare the difference between virtual plain CT images, enhanced CT images and actual plain CT images. Methods CT images of 50 patients underwent both plain and enhanced CT scanning in the same examination were included, and the volume CT dose index (CTDIvol) were recorded. Taken 40 cases as training set, the enhanced CT images were inputted, and the corresponding plain CT images were outputted to train U-Net neural network. Taken the rest 10 cases as test set, virtual plain CT images were generated through the trained U-Net. Then the virtual plain CT images were compared with actual plain CT images and enhanced CT images, and the parameters were also compared. Results The mean CTDIvol of 50 cases for plain CT scanning was (11.67±0.51)mGy, while for enhanced CT was (13.46±0.76)mGy. Virtual plain CT could reduce the average radiation dose by 46.44%. The mean absolute error (MAE) between enhanced and plain scan CT was (32.28±2.64)HU, and the structural similarity (SSIM) was 0.82±0.05. MAE between virtual plain CT and actual plain CT images was (6.72±1.31)HU, and SSIM was 0.98±0.02. There was no significant difference of the mean CT value of major organs between virtual plain CT and actual plain CT images (all P>0.05). Conclusion Based on U-Net neural network, virtual plain CT images generated by enhanced CT were consistent with actual plain scan CT images, hence reducing times of scanning and radiation dose.