【摘要】 目的：为了消除宫颈CT图像中存在的金属伪影，提出一种利用卷积神经网络（CNN）去除金属伪影的策略。方法：首先通过数值仿真得到金属伪影图像与目标图像（无伪影图像），构造训练测试数据集，利用含金属伪影的宫颈CT图像和对应的无伪影图像训练已搭建的CNN，进而得到去除宫颈CT图像金属伪影的CNN模型。结果：训练网络之前金属伪影图像与目标图像峰值信噪比（PSNR）平均值为26.0980dB。不同尺寸（25×25、50×50、100×100）的图像块训练网络得到去除金属伪影的图像与目标图像PSNR平均值分别为34.6079、38.3751、38.1838dB。结论：通过对仿真数据和临床数据进行实验，研究结果表明，本文方法能够快速有效地消除宫颈CT图像中的金属伪影，并且可以保留完整的组织结构信息。更多还原

【Abstract】 Objective To reduce metal artifacts in cervical CT images using convolutional neural network. Methods The metal artifact images and the target images(artifact-free images) were generated using numerical simulation for constructing training and test data sets. The cervical CT images with metal artifacts and paired cervical CT images without metal artifacts were input into the constructed convolutional neural network for training, and then a convolutional neural network model for metal artifact reduction in cervical CT images was obtained. Results Before network training, the average peak signal-tonoise ratio(PSNR) of the metal artifact images and the target images was 26.098 0 dB. The average PSNR of the metal artifact reduction images and the target images obtained by the training network trained by image patches of different sizes(25×25,50×50, 100×100) was 34.607 9, 38.375 1, and 38.183 8 dB, respectively. Conclusion Through experiments on simulation data and clinical data, it is revealed that the proposed method can effectively reduce metal artifacts and can retain relatively complete tissue texture information in cervical CT images.更多还原