基于MRI影像组学特征的机器学习模型预测胰腺癌CD_(8)^(+)T细胞浸润及预后的研究  

MRI radiomics-based machine learning model for predicting tumor-infiltrating CD_(8)^(+)T cells and prognosis of patients with pancreatic cancer

作　　者：卢明智 刘芳 方旭 边云 邵成伟 陆建平 李晶 Lu Mingzhi;Liu Fang;Fang Xu;Bian Yun;Shao Chengwei;Lu Jianping;Li Jing(Department of Radiotherapy,First Affiliated Hospital of Naval Medical University,Shanghai 200433,China;Department of Radiology,First Affiliated Hospital of Naval Medical University,Shanghai 200433,China)

机构地区：[1]海军军医大学第一附属医院放射治疗科,上海200433 [2]海军军医大学第一附属医院放射诊断科,上海200433

出　　处：《中华胰腺病杂志》2023年第5期344-352,共9页Chinese Journal of Pancreatology

基　　金：国家自然科学基金(81871352、82171915、82171930、82271972、82202125);申康三年行动计划重大临床研究项目(SHDC2020CR4073);上海市科技创新行动计划医学创新研究项目(21Y11910300、21ZR1478500)。

摘　　要：目的探讨基于MRI影像组学特征的机器学习模型预测胰腺癌CD_(8)^(+)T细胞浸润及患者预后的价值。方法回顾性分析2017年1月至2018年4月间海军军医大学第一附属医院156例术前7 d内接受MRI检查且术后经病理证实为胰腺癌患者的临床资料。依据国际预测模型建模共识,将2017年1月至12月共116例患者纳入训练集,2018年1月至4月共40例纳入验证集。以患者总生存期为结局变量,使用X-Tile软件获取CD_(8)^(+)T细胞占比的截点值,以此为界值,将患者分为高CD_(8)^(+)T细胞组和低CD_(8)^(+)T细胞组,比较两组的临床、病理学和影像学特征。使用3D Slicer软件对每例患者MRI平扫的T_(1)加权、T_(2)加权以及三期动态增强原始横断面图像逐层勾画感兴趣区进行肿瘤分割,使用Python程序包提取分割后的胰腺肿瘤影像组学特征,采用最小绝对收缩和选择算子(Lasso)回归方法对提取到的组学特征进行降维和选择,根据Lasso回归方程公式计算患者的组学分值。然后采用极端梯度提升(XGBoost)建立机器学习预测模型。绘制受试者工作特征曲线(ROC),计算曲线下面积(AUC)、灵敏度、特异度、准确度、阳性预测值和阴性预测值,评估模型的性能。结果X-Tile软件获取的CD_(8)^(+)T细胞含量的截点值为19.09%。高CD_(8)^(+)T细胞组患者中位生存时间比低CD_(8)^(+)T细胞组患者更长(25.51个月比22.92个月,P=0.007);两组在训练集的T分期和验证集的MRI图像测量的肿瘤大小方面差异有统计学意义(P值均<0.05)。共获得MRI图像1409个组学特征,经Lasso回归降维后得到19个与CD_(8)^(+)T细胞含量相关的组学特征。高CD_(8)^(+)T细胞组的组学分值为-0.43(-1.55~0.65),低CD_(8)^(+)T细胞组的组学分值为0.22(-0.68~2.54),两组间差异有统计学意义(P<0.001)。将肿瘤大小和组学分值纳入机器学习模型,模型在训练集的AUC值为0.90(95%CI 0.85~0.95),灵敏度、特异度、准确度、阳性Objective To investigate the value of machine learning model based on MRI in predicting the abundance of tumor infiltrating CD_(8)^(+)T cell and prognosis of pancreatic cancer patients.Methods The clinical data of 156 patients with pathological confirmed pancreatic cancer who underwent pre-operative MRI within 7 days before surgery in the First Affiliated Hospital of Naval Medical University from January 2017 to April 2018 was retrospectively analyzed.According to the international consensus on the predictive model,a total of 116 patients from January to December 2017 were included in the training set,and a total of 40 patients from January to April 2018 were included in the validation set.With the overall survival of patients as the outcome variable,X-Tile software was used to obtain cut-off values of the percentage of CD_(8)^(+)T cells,and all patients were divided into CD_(8)^(+)T-high and-low groups.The clinical,pathological and radiological features were compared between two groups.3D slicer software was used to draw the region of interest in each layer of the primary MR T_(1)-and T_(2)-weighted imaging,arterial phase,portal venous phase,and delayed phase images for tumor segmentation.Python package was applied to extract the radiomics features of pancreatic tumors after segmentation and the extracted features were reduced and chosen using the least absolute shrinkage and selection operator(Lasso)logistic regression algorithm.Lasso logistic regression formula was applied to calculate the rad-score.The extreme gradient boosting(XGBoost)were used to construct the machine learning predicted model.The models′performances were determined by area under the ROC curve(AUC),sensitivity,specificity,accuracy,positive predictive value,and negative predictive value.Results The cut-off value of the CD_(8)^(+)T-cell level was 19.09%as determined by the X-tile program.Patients in the high CD_(8)^(+)T cell group had a longer median survival than those in the low CD_(8)^(+)T cell group(25.51 month vs 22.92 month,P=0.007).Th

关 键 词：胰腺癌 CD_(8)阳性T淋巴细胞 磁共振成像 影像组学 预后 

分 类 号：R445.2[医药卫生—影像医学与核医学] R735.9[医药卫生—诊断学]