进展期胃癌生存预测:基于增强CT深度学习模型的构建
The value of deep learning in the construction of survival prediction models for advanced gastric cancer based on enhanced CT
张文娟 1张利文 2邓娟 1任铁柱 1徐敏 1周俊林1
作者信息
- 1. 730030 甘肃兰州,兰州大学第二医院放射科/甘肃省医学影像重点实验室/医学影像人工智能甘肃省国际科技合作基地
- 2. 100190 北京,中国科学院自动化所分子影像重点实验室
- 折叠
摘要
目的:探讨基于术前增强CT构建的深度学习(DL)模型对进展期胃癌(AGC)1、2、3年生存概率的预测价值.方法:回顾性分析2013年1月-2015年12月在本院经病理证实为AGC的337例患者的临床和CT资料.按照7∶3的比例将患者随机分为训练集(n=237)和验证集(n=100).采用数据增强技术增加训练集的数据量,随后基于术前C T增强静脉期图像构建残差卷积神经网络结构的DL模型,预测AGC患者1、2、3年的生存概率.经Cox单因素及多因素分析构建临床模型,然后联合DL模型和临床模型构建综合模型并绘制其诺莫图.计算各模型的Harrel 一致性指数(C-index)和风险比(HR),并应用Kaplan-Meier曲线、校准曲线及临床决策曲线比较3种模型对OS的预测效能.结果:在训练集和验证集中,临床模型、DL模型和综合模型的C-index值分别为0.70(95％CI:0.65～0.75)、0.72(95％CI:0.67～0.76)、0.74(95％CI:0.69～0.78)和 0.64(95％CI:0.56～0.71)、0.66(95％CI:0.58～0.73)、0.67(95％CI:0.59～0.74),表明综合模型具有最优的生存期预测能力;三个模型的HR分别为 2.72(95％CI:2.06～4.02)、2.88(95％CI:1.89～4.39)、2.72(95％CI:2.13～3.49)和 2.11(95％CI:1.43～3.11)、4.32(95％CI:1.66～11.24)、1.89(95％CI:1.36～2.60),均以 DL 模型的 HR 最高,表明 DL模型预测的高危人群具有更高的死亡风险.校准曲线分析显示基于综合模型的诺莫图预测AGC患者1、2、3年生存概率与实际的预后随访结果具有较高的一致性.临床决策曲线显示综合模型的净收益优于其它2种模型.结论:基于CT增强静脉期图像利用残差卷积神经网络构建的DL模型是一种良好的AGC患者生存风险评估模型,对AGC患者生存期的早期预判具有较高的临床应用价值.
Abstract
Objective:To explore the value of deep learning(DL)model based on preoperative enhanced CT to predict the 1-,2-and 3-year survival probability in patients with advanced gastric can-cer(AGC).Methods:From January 2013 to December 2015,the clinical and CT data of 337 patients with AGC confirmed by pathology in our hospital were retrospectively analyzed.All subjects were di-vided into a training set(n=237)and an external validation set(n=100)according to a ratio of 7∶3.A DL model of residual convolutional neural network was constructed based on preoperative contrast-enhanced venous phase CT images to predict the 1-,2-and 3-year survival probability of AGC patients.Data enhancement technology was used to increase the data amount of training set.Univariate and multivariate Cox regression analysis methods were used to construct clinical models,and then DL model and clinical model were integrated to construct comprehensive prediction models and a corre-sponding nomogram was created.Harrel agreement index(C-index)and risk ratio(HR)were calcula-ted,and Kaplan-Meier curve,calibration curve and clinical decision curve were used to compare the prediction performance of the three models for OS.Results:In the training set,the C-index values of comprehensive model,DL model and clinical model were 0.74(95％CI:0.69～0.78),0.72(95％CI:0.67～0.76)and 0.70(95％CI:0.65～0.75),respectively;and in the validation set,they were 0.67(95％CI:0.59～0.74),0.66(95％CI:0.58～0.73)and 0.64(95％CI:0.56～0.71),respectively.The re-sults indicated that the comprehensive model had the best survival prediction ability.In the training set,the HRs of the three models were 2.72(95％CI:2.06～4.02),2.88(95％CI:1.89～4.39)and 2.72(95％CI:2.13～3.49);in the validation set,the HRs of the three models were 2.11(95％CI:1.43～3.11),4.32(95％CI:1.66～11.24)and 1.89(95％CI:1.36～2.60).The results showed that the DL model had the highest HR in both training and validation set,which indicated that the patients in high-risk group predicted by the DL model had a higher risk of death.Calibration curve analysis showed that the probability of 1-,2-and 3-years survial probability predicted by nomogram based on the compre-hensive prediction model was in good agreement with the actual prognostic follow-up results.The clini-cal decision curve also proved that the net benefit of the comprehensive model was better than that of the other two models.Conclusion:The DL model constructed based on residual convolutional neural network in this study is a good survival risk assessment model,which has good application value for realizing early prediction of survival probability in AGC patients.
关键词
进展期胃癌/体层摄影术,X线计算机/残差卷积神经网/深度学习/预后Key words
Advanced gastric cancer/Tomography,X-ray computed/Residual convolutional neural network/Deep learning/Prognosis引用本文复制引用
基金项目
国家自然科学基金资助项目(82102151)
兰州市青年科技人才创新项目(2023-2-44)
甘肃省省级人才项目青年个人项目(甘组通字[2024]4号)
出版年
2024
NETL
NSTL
万方数据