基于深度学习的CT影像预测胃癌HER2状态  

Prediction of HER2 status in gastric cancer from CTimages based on deep learning

作　　者：张昊辰 邱炳江 崔艳芬 陈鑫[4,5] 姚丽莎 刘再毅 ZHANG Haochen;QIU Bingjiang;CUI Yanfen;CHEN Xin;YAO Lisha;LIU Zaiyi(School of Medicine,South China University of Technology,Guangzhou 510006,Guangdong,China;Guangdong Provincial People's Hospital(Guangdong Academy of Medical Sciences),Southern Medical University,Guangzhou 510080,China;Guangdong Cardiovascular Institute,Guangdong Provincial People′s Hospital,Guangdong Academy of Sciences,Guangzhou 510080,China;Department of Radiology,Guangzhou First People′s Hospital,Guangzhou 510180,Guangdong,China;Guangdong Provincial Key Laboratory of Artificial Intelligence in Medical Image Analysis and Application,Guangdong Provincial People′s Hospital,Guangdong Academy of Medical Sciences,Guangzhou 510080,Guangdong,China)

机构地区：[1]华南理工大学医学院,广东广州510006 [2]南方医科大学附属广东省人民医院(广东省医学科学院)放射科,广东广州510080 [3]广东省心血管病研究所,广东省人民医院,广东省医学科学院,广东广州510080 [4]广州市第一人民医院放射科,广东广州510180 [5]广东省医学影像智能分析与应用重点实验室,广东广州510080

出　　处：《暨南大学学报（自然科学与医学版）》2023年第1期78-86,共9页Journal of Jinan University(Natural Science & Medicine Edition)

基　　金：国家重点研发计划项目(2021YFF1201003);国家青年科学基金项目(82202142);中国博士后科学基金项目(2021M700897,2022M720857);广东省重点实验室项目(2022B1212010011)。

摘　　要：目的:构建并验证一个深度学习模型,旨在全自动流程地对胃癌病灶进行检出并预测胃癌患者人表皮生长因子受体(HER2)状态。方法:回顾性收集经手术病理证实为胃癌及明确HER2状态的135例胃癌患者CT及临床数据(HER2阴性88例;HER2阳性47例),随机分为训练集(n=95)和验证集(n=40)。基于级联nnUNet神经网络建立胃分割模型和胃癌自动检出模型,选择合适区域,以此建立基于ResNet网络的HER2状态预测模型,并验证模型的预测性能。结果:胃分割模型五折交叉验证的Dice系数为91.4%。胃癌检出模型在训练集和验证集中检出率分别为96.8%和97.5%。HER2预测模型敏感度训练集为78.1%,95%置信区间(CI)(61.2%,88.9%);验证集为71.4%,95%CI (45.4%,88.3%)。特异性训练集为67.9%,95%CI (54.8%,78.6%);验证集为69.2%,95%CI (50.0%,83.5%)。ROC曲线下面积(AUC)训练集为84.0%,95%CI (76.4%,90.7%);验证集为78.6%,95%CI (65.9%,90.0%)。结论:本研究建立的胃癌HER2预测模型不仅可自动检出胃癌,还作为非侵入性工具在预测HER2状态方面具有良好性能,具有指导临床评估的价值。Objective:To develop and validate a deep learning-based fully automated pipeline for gastric cancer detection and prediction of human epidermal growth receptor(HER2)status.Methods:A total of 135 gastric cancer cases with pathologically confirmed HER2 status(HER2 negative:n=88;HER2 positive:n=47)were retrospectively collected and randomly divided into training dataset(n=95)and validation dataset(n=40).A cascaded nnUNet was applied to establish an automatic detection model for gastric cancer.An appropriate region was selected by the detection model to develop a HER2 status prediction model based on the ResNet.The prediction model was verified by the validation dataset.Results:The average Dice coefficient of 91.4%for the gastric segmentation model was obtained by 5-fold cross-validation.The detection rates of the gastric cancer detection model in the training set of 95 cases and the validation set of 40 cases were 96.8%and 97.5%,respectively.The sensitivity of the HER2 status prediction model in the training set was 78.1%,95%confidence interval(CI)(61.2%,88.9%).The sensitivity in the validation set was 71.4%,95%CI(45.4%,88.3%).The specificity in the training set was 67.9%,95%CI:(54.8%,78.6%).The specificity in the validation set was 69.2%,95%CI(50.0%,83.5%).The area under the curve(AUC)of the HER2 status prediction model in the training set was 84.0%,95%CI(76.4%,90.7%).The AUC in the validation set was 78.6%,95%CI(65.9%,90.0%).Conclusion:The gastric cancer HER2 status prediction model established in this study can fully automatically detect and localize gastric cancer,and it achieve good performance in HER2 prediction,and guide clinical treatment as a non-invasive tool.

关 键 词：胃癌 深度学习 人表皮生长因子受体2(HER2) 电子计算机断层摄影(CT) 

分 类 号：R735.2[医药卫生—肿瘤] R318[医药卫生—临床医学]