基于术前^(18)F-FDG PET/CT影像组学机器学习模型预测Ⅱ~Ⅲ期结直肠腺癌病理高危状态及预后评估  

Machine Learning Models Based on Preoperative ~(18)F-FDG PET/CT Radiomics Features for Predicting Pathological High-Risk Status and Prognostic Evaluation in Stage Ⅱ-Ⅲ Colorectal Adenocarcinoma

作　　者：王冰[1] 屈伟[1] 郑龙[1] 胡添源 张占文[3] WANG Bing;QU Wei;ZHENG Long(Department of Nuclear Medicine,The Second Affiliated Hospital of Xi'an Jiaotong University,Xi'an,Shaanxi Province 710004,P.R.China)

机构地区：[1]西安交通大学第二附属医院核医学科,710004 [2]西安市红会医院核医学科,710054 [3]中山大学附属第六医院核医学科,广州510655

出　　处：《临床放射学杂志》2025年第4期670-677,共8页Journal of Clinical Radiology

基　　金：广东省自然科学基金-面上项目基金资助项目(编号:2023A1515011300)。

摘　　要：目的 基于氟代脱氧葡萄糖正电子发射断层成像/X线计算机体层成像(^(18)F-FDG PET/CT)影像组学特征,构建机器学习模型用于术前预测结直肠腺癌患者的病理高危状态[淋巴血管浸润(LVI)、周围神经侵犯(PNI)、病理T4期或低分化],并评估其对患者预后的价值。方法 回顾性分析2011年7月至2021年7月接受术前^(18)F-FDG PET/CT检查并确诊为Ⅱ~Ⅲ期结直肠腺癌的332例患者,并按照8∶2随机划分为训练集和测试集。分别提取PET和CT图像的影像组学特征,经过最大相关最小冗余(mRMR)、最小绝对收缩与选择算子(LASSO)算法等筛选和降维保留最优特征,并通过Logistic回归(LR)、支持向量机(SVM)、k近邻(KNN)、随机森林(RF)、极端梯度提升(XGBoost)及轻量级梯度提升机(LightGBM)6种机器学习算法分别构建全集及Ⅲ期患者预测模型。绘制受试者工作特征曲线(ROC)及决策曲线分析(DCA),计算曲线下面积(AUC)、准确率、敏感度、特异度用以评价模型性能。结果 在全集患者模型中,LR算法表现最优,训练集AUC为0.755(95%CI:0.696~0.813),准确率为0.714,测试集AUC为0.758(95%CI:0.638~0.879),准确率为0.758。在Ⅲ期患者模型中,RF算法性能最佳,训练集AUC为0.943(95%CI:0.905~0.981),准确率为0.863,测试集AUC为0.885(95%CI:0.742~1.000),准确率为0.828。DCA表明2种最优模型在预测病理高危状态及优化临床决策方面具有优异的应用价值。Kaplan-Meier生存曲线分析显示,病理高危状态显著增加患者复发转移风险,在Ⅲ期患者中尤为显著。结论 基于术前^(18)F-FDG PET/CT影像组学特征的机器学习模型能够较准确预测结直肠腺癌患者的病理高危状态,尤其在Ⅲ期患者中表现出更优的诊断效能,为术前个体化治疗策略制定和改善患者预后提供了参考依据。Objective This study aimed to develop machine learning models based on radiomics signatures derived from ~(18)F-fluorodeoxyglucose positron emission tomography/computed tomography(~(18)F-FDG PET/CT) to predict pathological high-risk status(lymphovascular invasion,perineural invasion,pathological T4 stage,or poor differentiation) in patients with stage Ⅱ-Ⅲ colorectal adenocarcinoma and to assess its impact on prognosis.Methods A retrospective analysis was performed on 332 cases of stage Ⅱ-Ⅲ colorectal adenocarcinoma diagnosed between July 2011 and July 2021 at The Sixth Affiliated Hospital,Sun Yat-sen University.The patients were randomly allocated into a training cohort and a test cohort in a ratio of 8∶2.Radiomic features were extracted separately from PET and CT images,and optimal features were selected through the maximum correlation minimum redundancy(mRMR) and least absolute shrinkage and selection(LASSO).Predictive models were built using six machine learning algorithms,including logistic regression(LR),support vector machine(SVM),k-nearest neighbor(KNN),random forest(RF),extreme gradient boosting(XGBoost),and light gradient boosting machine(LightGBM).Models were developed for the overall cohort and separately for stage Ⅲ cases.Receiver operating characteristic(ROC) curves and decision curve analysis(DCA) were generated,and the area under the curve(AUC),accuracy,sensitivity,and specificity were calculated to evaluate model performance.Results For the overall cohort,the LR model showed the best performance,with an AUC of 0.755(95%CI:0.696-0.813) and accuracy of 0.714 in the training set,and an AUC of 0.758(95%CI:0.638-0.879) and accuracy of 0.758 in the test set.Among stage Ⅲ patients,the RF model outperformed other algorithms,achieving an AUC of 0.943(95%CI:0.905-0.981) and accuracy of 0.863 in the training set,and an AUC of 0.885(95%CI:0.742-1.000) and accuracy of 0.828 in the test set.DCA demonstrated significant clinical utility for both optimal models in predicting pathological high-risk s

关 键 词：正电子发射断层成像 影像组学 机器学习 结直肠腺癌 病理高危状态 

分 类 号：R73[医药卫生—肿瘤]