【摘要】 目的基于CT影像建立手工特征和深度学习特征的标签,并联合临床危险因素建立影像组学模型,探讨模型对术前胃癌病人Ki-67表达水平的诊断效能。方法回顾性收集2009年1月—2019年1月行上腹部CT增强扫描且术后行胃癌组织Ki-67表达水平检测的468例胃癌病人的影像和临床病理资料,随机分为训练集(310例)和验证集(158例),并根据Ki-67表达水平将病人分为高表达组(训练集177例,验证集79例)和低表达组(训练集133例,验证集79例)。在训练集中基于CT影像分别提取、筛选肿瘤的手工特征及深度学习特征以建立影像组学标签,并联合临床信息进行多因素逻辑回归分析,构建可术前个体化预测胃癌Ki-67表达水平的影像组学模型。采用受试者操作特征(ROC)曲线分别评价训练集与验证集中影像组学标签及联合模型预测Ki-67表达水平的效能,并计算曲线下面积(AUC)。采用校准曲线评估联合预测模型术前预测Ki-67状态的结果与术后病理真实状态的拟合度,采用决策曲线分析(DCA)计算联合模型的净获益的阈值概率。采用独立样本t检验或卡方检验对高低表达组间的临床和病理特征、影像组学评分进行比较。结果最终获得与胃癌Ki-67水平显著相关的20个影像组学特征(9个手工特征,11个深度学习特征),结合多因素逻辑回归分析得到的临床危险因素(年龄)构建影像-临床联合预测模型,并将该模型可视化为诺谟图。ROC曲线示单纯的影像组学标签预测Ki-67表达水平的AUC在训练集和验证集分别为0.637(95%CI:0.570～0.704)和0.724(95%CI:0.641～0.807)。加入了临床危险因素的联合模型可在术前更好地预测胃癌Ki-67表达水平,训练集和验证集的AUC分别为0.656(95%CI:0.589～0.724)和0.733(95%CI:0.650～0.816)。校准曲线评估显示联合模型在验证集中有更好的拟合度,DCA表明影像组学联合模型具有良好的临床应用价值。结论基于CT影像建立的手工和深度影像组学标签,并结合临床危险因素建立的联合预测模型可作为术前评估胃癌病人Ki-67表达状态的一种无创性辅助工具,有利于协助临床决策,改善病人预后。更多还原

【Abstract】 Objective To assess the performance of the radiomics model which is established from the combination of clinical risk factors, hand-crafted features and deep learning features extracted from CT images for preoperatively predicting the Ki-67 status in patients with gastric cancer. Methods The imaging and clinicopathological data of four hundred and sixty-eight gastric cancer patients who underwent contrast-enhanced CT scan of the upper abdomen and postoperative Ki-67 measurement from January 2009 to January 2019, were retrospectively analyzed. The enrolled patients were randomly divided into the training set(310 patients) and validation set(158 patients), and divided into high and low expression group according to the expression level of Ki-67(training and validation set:177 and 133 patients, 79 and 79 patients, respectively). Based on the CT images of training set, hand-crafted features and deep learning features were respectively extracted and selected. Then a radiomics signature was developed. Next, multivariate logistic regression analysis was carried out in combination with clinical information to build a radiomics model that can individually predict the Ki-67 expression level of gastric cancer before surgery. The predictive performances of the radiomics signatures and combined model for Ki-67 status were respectively evaluated with receiver operating characteristic(ROC) curve in training and validation set, and the areas under the curve(AUC) were calculated. The calibration curve was used to evaluate the degree of fitting between the predicted Ki-67 status of the combined prediction model and the postoperative pathological reality.Decision curve analysis(DCA) was used to evaluate the threshold probability of net benefit of combined model. The independent samples t test or chi-square test was used to compare the clinicopathologic information and Rad-score between the high and low expression groups. Results Twenty selected radiomics features(9 handcrafted features, and 11 deep learning features) were significantly associated with the Ki-67 status in gastric cancer. The radiomics signature was incorporated with the clinical risk factors(age) obtained by multivariate logistic regression analysis to build an imaging combined clinical prediction model, which was then visualized as a nomogram. The ROC curve showed that the AUC for predicting Ki-67 level with the radiomics only were 0.637(95%CI: 0.570-0.704) and 0.724(95%CI: 0.641-0.807) in the training and validation sets, respectively. The combined model with clinical risk factors improved the prediction of the Ki-67 level before surgery, and achieved the AUC of 0.656(95%CI: 0.589-0.724) and 0.733(95%CI: 0.650-0.816) in the training and validation sets, respectively. The calibration curves showed that the combined model has a better fit in the validation set than in the training set. The DCA showed the radiomics nomogram was clinically useful. Conclusion This study presents a combined model that incorporates the hand-crafted, deep learning radiomics signature based on CT images and clinical risk factors, which can be used for individualized preoperative prediction of Ki-67 status in patients with gastric cancer, and could serve as a noninvasive tool to assist clinical decision-making and improve patient’s outcome.更多还原