基于3D深度卷积神经网络依据MRI生成伪CT的研究

Study on generating pseudo-CT image based on MRI using 3D deep convolutional neural network

目的:研究一种依据MRI生成伪CT的方法,从而减少放疗过程中额外CT的使用,降低患者辐射剂量,提高放疗精准度。方法:提出一种基于3D深度卷积神经网络(DCNN)的预测算法,利用单张图像的解剖特征以及相邻图像层之间的关联信息,从而提高了图像特征提取的准确性。采用U-net网络结构,通过编码部分的卷积层、池化层和解码部分的上采样、卷积层,对MRI和对应的CT进行端到端转换的学习。采集13例患者图像数据,应用留一交叉验证的方法,分别对3D DCNN和2D DCNN的伪CT结果与原始CT进行对照比较。结果:提出的3D DCNN算法的平均绝对误差(MAE)为86 HU,远小于2D DCNN的136 HU。结论:3D DCNN算法能更准确的生成伪CT,明显改善了骨骼、空气与软组织之间的误转化。

Objective:To study a method of generating pseudo-CT based on MRI,for reducing the use of extra CT in the course of radiotherapy,reducing the radiation dose of patients and improving the accuracy of radiotherapy.Methods:A prediction algorithm based on 3D deep convolutional neural network(DCNN)was proposed.The accuracy of image feature extraction was improved by utilizing the anatomical features of a single image and the correlation information between adjacent image slices.Using U-net network structure,the end-to-end conversion from MRI to the corresponding CT was studied by convolution layers and pooling layers in the encoding portion,and upsampling layers and convolution layers in the decoding portion.The image data of 13 patients were collected and tested.The results of generated pseudo-CT for 3D DCNN and 2D DCNN were compared with original CT using the method of leave-one-out cross validation,respectively.Results:A mean absolute error(MAE)of the proposed 3D DCNN algorithm was 86 HU,which was much smaller than the 136 HU of the 2D DCNN.Conclusion:The 3D DCNN algorithm can more accurately generate pseudo-CT and significantly improve the mis-conversion among bone,air and soft tissue.