基于CT深度学习模型预测局部进展期胃癌术后预后的研究  

Prognostic Value of CT-Based Deep Learning Model in Patients with Locally Advanced Gastric Cancer

作　　者：陈鑫[1] 张国琴 阮秀杭 魏新华[1] CHEN Xin;ZHANG Guoqin;RUAN Xiuhang(Department of Radiology,Guangzhou First People’s Hospital,Guangzhou,Guangdong Province 510180,P.R.China)

机构地区：[1]广州市第一人民医院放射科,510180

出　　处：《临床放射学杂志》2020年第7期1342-1346,共5页Journal of Clinical Radiology

基　　金：国家自然科学基金项目(No.81601469);广州市卫生健康科技项目(No.20191A011002);广州市科技计划项目(No.201804010032)资助。

摘　　要：目的探讨基于CT深度学习构建的模型在预测局部进展期胃癌患者术后预后中的价值方法回顾性搜集200例(训练组134例,验证组66例)术后病理确诊为局部进展期胃癌患者的术前CT增强图像和临床资料。在门静脉期CT图像肿瘤最大层面提取深度学习特征,用LASSO Cox回归方法选择特征并构建标签,然后通过多因素Cox回归模型融合标签和临床病理信息构建预测模型,并用诺莫图对模型可视化。采用区分度、校准度和临床决策曲线等评价模型的预测效能。结果最终筛选出10个深度学习特征构建了深度学习标签,标签在训练组和验证组中均与总体生存时间显著相关(P<0.001和P=0.010)。融合深度学习标签和肿瘤TNM分期构建的预测模型在训练组[C-index(95%CI)=0.776(0.718~0.833)]和验证组[C-index(95%CI)=0.797(0.680~0.914)]均有较好的区分度和校准度。决策曲线分析表明预测模型有较好的临床实用性。结论基于术前CT图像的深度学习模型可个体化预测局部进展期胃癌患者术后预后,有望辅助临床治疗决策。Objective To build a CT-based deep learning model so as to predict prognosis in locally advanced gastric cancer.Methods A total of 200 patients with pathologically confirmed locally advanced gastric cancer were retrospectively enrolled in this study and divided into training(n=134)and validation sets(n=66).Tumors were segmented and analyzed to extract deep learning features on the largest cross-sectional slice of portal venous phase CT images.After feature selection using LASSO Cox regression,deep learning signature was built with the selected key features.Prediction model was built by using multivariate Cox regression incorporating deep learning signature and clinico-pathological variables and then visualized with normogram.Performance of the model was then assessed with respect to discrimination,calibration and clinical usefulness.Results Ten deep learning features were selected in the training dataset to build the deep learning signature.The deep learning signature was associated with OS in the training and validation cohort(P<0.001 and P=0.010 for training and validation sets,respectively).Normogram integrating the signature and TNM stage resulted in good discrimination[C-index(95%confidence interval):0.776(0.718-0.833)for training dataset,0.797(0.680-0.914)for validation dataset].Decision curve analysis demonstrated the clinical usefulness of the combined normogram.Conclusion The deep learning model based on pre-operative CT images is an independent biomarker for individualized OS estimation in patients with locally advanced gastric cancer,which can aid clinical decision-making.

关 键 词：局部进展期胃癌 深度学习 预后预测 

分 类 号：R735.2[医药卫生—肿瘤] R730.44[医药卫生—临床医学]