目的:免疫检查点抑制剂(ICIs)已经彻底改变了非小细胞肺癌(NSCLC)的治疗,而程序性细胞死亡配体1(PD-L1)是一种伴随生物标志物。本研究旨在使用基线动脉期增强CT(APECT)构建有效的影像组学模型,以预测NSCLC中PD-L1的表达和免疫治疗的预后。
方法:我们从已发表的多中心临床试验中纳入的204名患者的基线APECT图像中提取了影像组学特征,该临床试验于2018年8月23日开始,并于2019年11月15日结束(ClinicalTrials.gov:NCT03607539)。在这些病人中,来自选定中心的146名患者被分配到训练队列。使用最小绝对收缩和选择算子(LASSO)方法来减少影像组学特征的维度并计算肿瘤评分。模型是使用朴素贝叶斯创建的,决策树,XGBoost,和根据肿瘤评分的随机森林算法。然后在包括来自其余中心的58名患者的独立验证队列中验证这些模型。
结果:随机森林算法优于其他方法。在三分类方案中,随机森林模型在训练和验证队列中实现曲线下面积(AUC)值为0.98和0.94,分别。在双分类方案中,在训练和验证队列中,随机森林模型的AUC分别为0.99(95CI:0.97-1.0,P<0.0001)和0.93(95CI:0.83-0.98,P<0.0001),分别。此外,通过该模型分类为PD-L1高表达的患者仅在验证性sintilimab组中可以预测治疗反应(AUC=0.859,95CI:0.7~0.96,P<0.001)和提高生存率(HR=0.2,95CI:0.08~0.53,P=0.001).
结论:基于APECT的放射组学模型代表了一种潜在的非侵入性方法,可以可靠地预测NSCLC患者的PD-L1表达和ICI治疗结果。这可以显着改善精确的癌症免疫疗法。
OBJECTIVE: Immune checkpoint inhibitors (ICIs) have revolutionized the treatment of non-small cell lung cancer (NSCLC) and programmed cell death-ligand 1 (PD-L1) is a companion biomarker. This study aims to use baseline arterial-phase enhanced CT (APECT) to construct efficient radiomic models for predicting PD-L1 expression and immunotherapy prognosis in NSCLC.
METHODS: We extracted radiomics features from the baseline APECT images of 204 patients enrolled in a published multicenter clinical trial that commenced on August 23, 2018, and concluded on November 15, 2019 (ClinicalTrials.gov: NCT03607539). Of these patients, 146 patients from selected centers were assigned to the training cohort. The least absolute shrinkage and selection operator (LASSO) method was used to reduce dimensionality of radiomics features and calculate tumor scores. Models were created using naive bayes, decision trees, XGBoost, and random forest algorithms according to tumor scores. These models were then validated in an independent validation cohort comprising 58 patients from the remaining centers.
RESULTS: The random forest algorithm outperformed the other methods. In the three-classification scenario, the random forest model achieving the area under the curve (AUC) values of 0.98 and 0.94 in the training and validation cohorts, respectively. In the two-classification scenario, the random forest model achieved AUCs of 0.99 (95%CI: 0.97-1.0, P < 0.0001) and 0.93 (95%CI: 0.83-0.98, P < 0.0001) in the training and validation cohorts, respectively. Furthermore, patients classified as PD-L1 high-expression by this model can predict treatment response (AUC=0.859, 95%CI: 0.7-0.96, P < 0.001) and improved survival (HR=0.2, 95%CI: 0.08-0.53, P = 0.001) only in validation sintilimab arm.
CONCLUSIONS: Radiomics models based on APECT represent a potential non-invasive approach to robustly predict PD-L1 expression and ICI treatment outcomes in patients with NSCLC, which could significantly improve precision cancer immunotherapy.