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 cancerbased on enhanced CT

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CT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。

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 divided 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 corresponding 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 results 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 comprehensive prediction model was in good agreement with the actual prognostic follow-up results.The clinical 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.