基于靶扫描的3D卷积网络及基因检测在肺癌筛查中的运用  

Application of 3D convolution network and gene detection based on target scanning in lung cancer screening

作　　者：徐存来 曹卓[1,2] 尹章勇 陈璇 李雨玲 龚易莎 蒋奕薇 潘炯伟 XU Cunlai;CAO Zhuo;YIN Zhangyong;CHEN Xuan;LI Yuling;GONG Yisha;JIANG Yiwei;PAN Jiongwei(Department of Respiratory and Critical Care,Lishui People's Hospital f Lishui 323000,China;The First Clinical Medical College of Wenzhou Medical University,Wenzhou 325035,China)

机构地区：[1]丽水市人民医院呼吸与危重症科,浙江丽水323000 [2]温州医科大学第一临床医学院,浙江温州325035

出　　处：《健康研究》2022年第6期636-638,642,共4页Health Research

基　　金：丽水市科技局重点研发计划(2019ZDYF20);浙江省医药卫生计划项目(2021KY410,2023RC314)。

摘　　要：目的 创建基于靶扫描的3D卷积神经网络(3D-CNN)联合基因检测的早期肺癌筛查模型。方法 以80例Ⅰ/Ⅱ期肺癌患者(肺癌组)和80例确诊良性肺结节者(良性组)为研究对象,检测其外周血pl6、RASSFIA基因启动子甲基化水平,并收集患者初次就诊时的肺结节靶扫描+二三维重建图像,共同建立早期肺癌筛查模型,与Fisher及决策树模型比较其诊断价值。结果 肺癌组和良性组外周血pl6、RASSFIA基因启动子甲基化水平比较,差异均有统计学意义(P=0.008、0.038)。采用3D-CNN网络模型后,预测集准确率(83.33%)比训练集准确率(72.00%)提高,但二者差异无统计学意义(χ^(2)=0.602,P=0.438)。与Fisher、决策树预测模型比较,3D-CNN模式对早期肺癌的诊断灵敏度(79.52%)、特异度(89.24%)最高。结论 基于靶扫描的3D-CNN联合基因检测模型可用于早期肺癌筛查。Objective To create a 3 D convolutional neural network(3 D-CNN) combined with gene detection model for early lung cancer screening based on target scanning. Methods Eighty patients with stage Ⅰ/Ⅱ lung cancer(lung cancer group) and 80 patients with diagnosed benign lung nodules(benign group) were selected as the study subjects. The methylation levels of pl6 and RASSFIA gene promoter regions in peripheral blood were detected, and the lung nodule target scan + two-dimensional and three-dimensional reconstruction images at the first visit of patients were collected. A screening model for early lung cancer was established, and its diagnostic value was compared with Fisher and decision tree models. Results The methylation levels of pl6 and RASSFIA gene promoter in peripheral blood of lung cancer group and benign group were significantly different(P=0.008, 0.038). After using 3 D-CNN network model, the accuracy rate of prediction set(83.33%) is higher than that of training set(72.00%), but there is no significant difference between the two(χ^(2)=0.602, P=0.438). Compared with Fisher and decision tree prediction models, 3 D-CNN model has the highest diagnostic sensitivity(79.52%) and specificity(89.24%) for early lung cancer. Conclusions 3 D-CNN combined gene detection model based on target scanning can be used for early lung cancer screening.

关 键 词：肺癌筛查 靶扫描 卷积神经网络 DNA甲基化 智慧健康 

分 类 号：R5[医药卫生—内科学]