基于ADC图的三维卷积神经网络模型在判断直肠癌T分期的应用  

Application of 3D Convolutional Neural Network Model Based on ADC Map in Determining T-Stage of Rectal Cancer

作　　者：王昌盛 方哲明[1,2] 陈德华[1,2] 郭飞宝[3] 陈君 林晓君[1,2] 郭翌 WANG Changsheng;FANG Zheming;CHEN Dehua;GUO Feibao;CHEN Jun;LIN Xiaojun;GUO Yi(Department of Imaging,The First Affiliated Hospital of Fujian Medical University,Fuzhou 350005,China;Department of Imaging,National Regional Medical Center,Binhai Campus of The First Affiliated Hospital of Fujian Medical University,Fuzhou 350212,China;Department of Radiotherapy,Cancer Center,The First Affiliated Hospital of Fujian Medical University,Fuzhou 350005,China)

机构地区：[1]福建医科大学附属第一医院影像科,福州350005 [2]福建医科大学附属第一医院滨海院区国家区域医疗中心影像科,福州350212 [3]福建医科大学附属第一医院肿瘤中心放疗科,福州350005

出　　处：《福建医科大学学报》2023年第1期41-45,共5页Journal of Fujian Medical University

基　　金：福建省自然科学基金项目(2021J05146)。

摘　　要：目的利用三维卷积神经网络与磁共振弥散加权成像(DWI)序列的ADC图对直肠癌患者肿瘤T分期进行分类判断,提高分期准确度。方法回顾性分析183例直肠癌病例,其中训练集160例,测试集23例(T1/T2期13例,T3/T4期10例)。训练集图像采用水平、垂直翻转等方式进行4倍扩充。基于三维卷积神经网络进行训练,采用十折交叉验证方法降低模型过拟合程度。根据测试结果数据绘制受试者工作特征(ROC)曲线,并计算曲线下面积(AUC),分析卷积网络模型的准确性与可靠性。结果测试集测试结果显示,卷积网络模型判断肿瘤T分期的准确率为82.6%,ROC曲线的AUC为0.850,敏感度和特异度分别为84.6%和80.0%。结论基于卷积神经网络模型与ADC图自动判断直肠癌肿瘤T分期相比人工分期提高了准确性与效率。Objective To improve staging accuracy for rectal cancer by using a three-dimensional(3D)convolutional neural network and the ADC values of DWI images.Methods A retrospective analysis of 183 cases of rectal cancer was performed.160 cases were used as training data and 23 cases as test data(T1/T2:13,T3/T4:10),and the training data were expanded 4 times by horizontal and vertical image flipping.The tumor T-stage was classified and predicted based on the 3D convolutional neural network,and the 10-fold cross-validation method was used to reduce over-fitting.The diagnosis accuracy and reliability of the neural network model were analyzed using ROC curve.Results 23 cases were tested using our model.The diagnosis accuracy of the T-stage was 82.6%,and the AUC value,sensitivity,and specificity were 0.850,84.6%,and 80.0%,respectively.Conclusion A 3D convolutional neural network model based on ADC values without pre-extracting features was proposed to predict the T-stage of rectal cancer,which improved the accuracy and efficiency compared with the manual staging method.

关 键 词：卷积神经网络 ADC图 直肠癌 T分期 

分 类 号：R735.37[医药卫生—肿瘤]