基于深度学习的影像组学预测直肠癌T2与T3分期

Radiomics based on deep learning to predict T2 and T3 staging of rectal cancer

目的本研究探讨基于MRI轴位高分辨T2WI图像的深度学习(deep learning,DL)影像组学在术前预测直肠癌T2与T3期的价值。材料与方法回顾性分析皖南医学院第一附属医院(弋矶山医院)2018年1月至2022年12月361例经术后病理证实的T2与T3期直肠癌患者的完整资料,其中T2期100例,T3期261例,按7∶3采用分层抽样将患者随机分为训练集(n=262)与测试集(n=99)。采用单因素与多因素logistic回归分析筛选临床影像特征独立危险因素。利用ResNet-18模型作为DL特征提取的基础模型,分别基于轴位高分辨T2WI图像提取手工影像组学(hand-crafted radiomic,HCR)特征及DL影像组学特征,分别基于临床影像特征、HCR特征、DL特征及三者组合特征利用支持向量机(support vector machine,SVM)、K最近邻(K-nearest neighbor,KNN)、极端梯度增强机(extreme gradient boosting,XGBoost)三种算法构建12个机器学习模型,采用ROC曲线下面积(area under the curve,AUC)评价各模型的诊断效能,确定最优模型作为输出模型。结果单因素与多因素logistic回归分析临床影像特征中碳水化合物抗原(carbohydrate antigen 199,CA19-9)[95%置信区间(confidence interval,CI):1.150-1.820,P=0.002]及肿瘤长径(longest diameter,LD)(95%CI:1.159-22.584,P=0.031)为预测T2与T3期直肠癌的独立危险因素,构建的所有模型中组合特征模型效能均高于单独特征模型,训练集XGBoost分类器模型效能最高,AUC为0.998(95%CI:0.995-1.000),作为本研究输出模型。结论基于MRI轴位高分辨T2WI图像的DL影像组学机器学习模型可有效预测直肠癌T2与T3期,其中训练集组合特征的XGBoost分类器模型效能最佳。

Objective:To explore the value of deep learning(DL)imageomics based on MRI axial high-resolution T2WI images in predicting T2 and T3 stages of rectal cancer before surgery.Materials and Methods:Retrospective analysis of the complete data of 361 patients with T2 and T3 stage rectal cancer confirmed by postoperative pathology at the First Affiliated Hospital of Wannan Medical College(Yijishan Hospital)from January 2018 to December 2022.Among them,there were 100 cases in T2 phase and 261 cases in T3 phase.Patients were randomly divided into a training set(n=262)and a testing set(n=99)using stratified sampling at 7∶3.Univariate and multivariate logistic regression analysis was used to screen independent risk factors for clinical imaging features.The ResNet-18 model was employed as the foundational model for DL feature extraction.Subsequently,twelve machine learning models were developed by incorporating clinical imaging features,hand‑crafted radiomi features,DL features,and their combined features.The support vector machine(SVM),K-nearest neighbor(KNN),and extreme gradient enhancement machine(XGBoost)algorithms were utilized for constructing these models.The diagnostic performance of each model was assessed by calculating the area under the curve(AUC)of the subject.Finally,the model with the highest performance was identified as the optimal output model.Results:The results of both univariate and multivariate logistic regression analysis indicate that carbohydrate antigen(CA19-9)[95%confidence interval(CI):1.150-1.820,P=0.002]and tumor length(LD)(95%CI:1.159-22.584,P=0.031)were independent risk factors for predicting T2 and T3 stage rectal cancer based on clinical imaging features.Among all the models constructed,the performance of combined feature model was higher than that of individual feature model,and the training set XGBoost classifier model had the highest performance,with an AUC of 0.998(95%CI:0.995-1.000),and was therefore selected as the output model for this study.Conclusions:The DL radiomics machine learning model based on MRI axial high-resolution T2WI images can effectively predict T2 and T3 stages of rectal cancer,with the XGBoost classifier model with combined features of the training set having the best performance.