摘要
目的:研究一种基于CT图像层间插值的方法,用于放射治疗过程中的患者摆位验证,从而提高放疗精度。方法:采用一种基于3D卷积和膨胀卷积神经网络(3D CNN-DCNN)算法,利用相邻图像层之间的关联信息重建中间层图像。采用U-Net网络架构,通过编码部分的卷积层、膨胀卷积层、池化层和解码部分的上采样层、卷积层、膨胀卷积层,对CT进行端到端的学习。采集20例患者图像数据,采用留一交叉验证的方法训练验证模型,分别对神经网络和线性插值的预测CT与原始薄层CT进行对照比较。结果:3D CNN-DCNN的平均绝对误差(MAE)为34 HU,远小于线性插值的55 HU。除此之外,骨骼的Dice相似系数(DSC)为0.95,大于线性插值方法的0.89。结论:与传统线性插值方法相比,3D CNN-DCNN算法可以更准确的重建薄层CT,明显改善了插值伪影、图像失真和锯齿状现象。
Purpose:To study a method based on interslice interpolation of CT images for the verification of patient setup during radiotherapy,and to improve the accuracy of radiotherapy.Methods:An algorithm based on 3D convolution and dilated convolutional neural network(3D CNN-DCNN)was used to reconstruct the intermediate slice images of CT by utilizing the correlation information between adjacent image slices.By using the U-Net network structure,end-to-end learning was done on CT images by convolutional layer,dilated convolutional layer,pooling layer in the encoding portion and up sampling layer,convolutional layer,and dilated convolutional layer in the decoding portion.The neural network was trained and tested with the image data of 20 patients using the method of leave-one-out cross validation.The predicted CT images from the neural network and linear interpolation were compared with the original thin-slice CT respectively.Results:The mean absolute error(MAE)of 3D CNN-DCNN is 34 HU,which was much smaller than that of linear interpolation(55 HU).The Dice similarity coefficient(DSC)of the bones was 0.95,which was larger than that of the linear interpolation method(0.89).Conclusion:Compared with linear interpolation method,the 3D CNN-DCNN algorithm can be used to reconstruct thin-slice CT more accurately,and improves the interpolation artifacts,image distortion and jaggedness.
作者
菅影超
马善达
王伟
JIAN Yingchao;MA Shanda;WANG Wei(R&D Department,Bejing Rayer Shiwei Medical Research Co Ltd.;R&D Department,Jiangou Rayer Medica!Technology Co.,Ltd.;Department of Radiotherapy,Tianjin Medical University Cancer Hospital)
出处
《中国医学计算机成像杂志》
CSCD
北大核心
2022年第6期669-675,共7页
Chinese Computed Medical Imaging
关键词
摆位验证
膨胀卷积神经网络
U-Net
留一交叉验证
平均绝对误差
Dice相似系数
Setup verification
Dilated convolutional neural network
U-Net
Leave-one-out cross validation
Mean absolute error
Dice similarity coefficient