摘要
目的:结合全卷积神经网络(Fully Convolutional Network,FCN)和多孔卷积(Atrous Convolution,AC)的深度学习方法,实现放射治疗计划图像的组织器官自动勾画。方法:选取122套已经由放疗医师勾画好正常器官结构轮廓的胸部患者CT图像,以其中71套图像(8 532张轴向切层图像)作为训练集,31套图像(5 559张轴向切层图像)作为验证集,20套图像(3 589张轴向切层图像)作为测试集。选取5种公开的FCN网络模型,并结合FCN和AC算法形成3种改进的深度卷积神经网络,即带孔全卷积神经网络(Dilation Fully Convolutional Network,D-FCN)。分别以训练集图像对上述8种网络进行调优训练,使用验证集图像在训练过程中对8种神经网络进行器官自动识别勾画验证,以获取各网络的最佳分割模型,最后使用测试集图像对充分训练后获取的最佳分割模型进行勾画测试,比较自动勾画与医师勾画的相似度系数(Dice)评价各模型的图像分割能力。结果:使用训练图像集进行充分调优训练后,实验的各个神经网络均表现出较好的自动图像分割能力,其中改进的D-FCN 4s网络模型在测试实验中具有最佳的自动分割效果,其全局Dice为94.38%,左肺、右肺、心包、气管和食道等单个结构自动勾画的Dice分别为96.49%、96.75%、86.27%、61.51%和65.63%。结论:提出了一种改进型全卷积神经网络D-FCN,实验测试表明该网络模型可以有效地提高胸部放疗计划图像的自动分割精度,并可同时进行多目标的自动分割。
Objective To combine fully convolutional network(FCN) and atrous convolution(AC) for realizing automation segmentation of tissues and organs in radiotherapy planning image. Methods A total of 122 sets of chest CT images were selected in this study, in which the normal organ structures were sketched by radiotherapy physician, including 71 sets of CT images(8 532 axial slice images) as training set, 31 sets of CT images(5 559 axial slice images) as validation set, and 20 sets of CT images(3 589 axial slice images) as test set. Five kinds of published FCN models were selected and combined with AC algorithm to form 3 kinds of improved deep convolutional neural networks, namely dilation fully convolutional network(D-FCN). Training set was used for fully fine-tuning the above 8 kinds of network, and validation set was applied to validate the automatic segmentation results for obtaining the optimal model of each network, and finally, test set was used to perform segmentation test for the optimal models. The Dice similarity coefficients of automatic segmentation and physician sketching were compared for evaluating the performances of these image segmentation models. Results After being fully fine-tuned with the use of training set, each neural network model showed good performances in automatic image segmentation. The improved D-FCN 4s model achieved the best automatic segmentation results in validation test, with a global Dice of 94.38%. The Dice of automatic segmentations of left lung, right lung, pericardium, trachea and esophagus was 96.49%, 96.75%, 86.27%, 61.51% and 65.63%,respectively. Conclusion An improved D-FCN is put forward in this study and the verification test shows that the improved DFCN can effectively improve the accuracy of automatic segmentation for radiotherapy planning image of chest, and segment multiple organs synchronously.
作者
邓金城
彭应林
刘常春
陈子杰
雷国胜
吴江华
张广顺
邓小武
DENG Jincheng;PENG Yinglin;LIU Changchun;CHEN Zijie;LEI Guosheng;WU Jianghua;ZHANG Guangshtm;DENG Xiaowu(Shenzhen Yino Intelligence Techonology Co, Ltd., Shenzhen 518057, China;Collaborative Innovation Center for Cancer Medicine/State Key Laboratory of Oncology in South China/Sun Yat-Sen University Cancer Center, Guangzhou 510060, China)
出处
《中国医学物理学杂志》
CSCD
2018年第6期621-627,共7页
Chinese Journal of Medical Physics
基金
国家重点研发计划(2017YFC0113200)
广东省科技计划项目(2015B020214002)
广州市科技计划项目(201508020105)
深圳市高技术产业化扶持计划项目(S2016I65100017)
关键词
深度学习
卷积神经网络
医学影像分割
相似度系数
放射治疗
deep learning
convolution neural network
medical imag
scrotation
similariyt coefficient
radiotherapy