The value of predicting cervical lymph node metastasis in papillary thyroid carcinoma based on multimo-dal radiomics combined with machine learning model
The value of predicting cervical lymph node metastasis in papillary thyroid carcinoma based on multimo-dal radiomics combined with machine learning model
郭建峰 1宋鑫洋 1沈天赐 1杜梦颖 1纪旭东 2杨峰1
扫码查看
点击上方二维码区域，可以放大扫码查看
作者信息
1. 441000 湖北,湖北医药学院附属襄阳市第一人民医院放射科
2. 441000 湖北,湖北医药学院附属襄阳市第一人民医院超声科
折叠
摘要
目的:构建基于超声、CT影像组学联合机器学习模型,探讨其预测甲状腺乳头状癌(PTC)颈部淋巴结转移(CLNM)的价值.方法:回顾性搜集198例经手术病理证实的PTC患者,其中CLNM组97例,无CLNM组101例,所有患者术前均接受甲状腺超声和增强CT扫描,按照7:3的比例随机分为训练集和测试集.采用单因素和多因素回归分析PTC患者发生CLNM的临床独立风险因素,构建临床风险模型.基于超声和CT图像提取并筛选最优影像组学特征,分别使用逻辑回归(LR)、支持向量机(SVM)、随机森林(RF)、极度随机树(ET)、K邻近算法(KNN)、XGBoost、监督学习集成模型(Light GBM)、前馈神经网络多层感知器(MLP)等8个机器学习模型构建PTC患者CLNM影像组学预测模型.绘制受试者工作特征(R O C)曲线评估模型效能,筛选出最优影像组学模型联合临床风险模型构建诺模图,应用决策曲线分析(DCA)对比不同模型的临床获益.结果:临床影像特征中年龄、肿瘤最大直径、超声报告淋巴结状态和CT报告淋巴结状态被筛选出来作为临床独立风险因素(P＜0.05).从超声和CT图像中分别筛选出14个和12个影像组学特征.8个机器学习模型中SVM模型表现最优,其预测CLNM状态的AUC在训练集为0.951(95％CI:0.920～0.982),测试集为0.857(95％CI:0.764～0.950).由临床模型和多模态影像组学模型构建的诺模图具有更好的预测效能,其预测CLNM 状态的 AUC 在训练集为 0.960(95％CI:0.933～0.987),测试集为 0.890(95％CI:0.812～0.968).DCA显示诺模图具有较高的临床净收益.结论:基于多模态影像组学联合机器学习模型可以更准确地预测PTC患者是否发生CLNM,具有一定的临床应用价值.
Abstract
Objective:To construct a combined machine learning model based on ultrasound(US)and computed tomography(CT)radiomics,and explore its predictive value for cervical lymph node metastasis(CLNM)in papillary thyroid carcinoma(PTC).Methods:One hundred and ninety-eight patients with PTC confirmed by surgery and pathology were retrospectively collected,including 97 patients in the CLNM group and 101 patients in the non-CLNM group.All patients underwent pre-operative thyroid US and contrast-enhanced CT scans,and were randomly divided into a training set and testing set in a 7:3 ratio.Univariate and multivariate regression analyses were performed to identi-fy independent clinical risk factors for CLNM in PTC,and a clinical risk model was constructed.Opti-mal radiomics features were extracted and selected from US and CT images.Eight machine learning models,including logistic regression(LR),support vector machine(SVM),random forest(RF),extra Trees(ET),K-nearest neighbor algorithm(KNN),XGBoost,light gradient boosting machine(Light GBM),and multi-layer perceptron(MLP),were used to construct a CLNM radiomics prediction model for PTC patients.Receiver operating characteristic(ROC)curves were generated to evaluate model performance.The optimal radiomics model combined with the clinical risk model was used to develop a nomogram,and decision curve analysis(DCA)was applied to compare the clinical benefits of different models.Results:Age,maximum tumor diameter,US-reported lymph node status,and CT-reported lymph node status were identified as independent clinical risk factors(P＜0.05).Fourteen and twelve radiomics features were selected from US and CT images,respectively.The SVM model performed best among the eight machine learning models,with an AUC of 0.951(95％CI:0.920～0.982)in the training set and an AUC of 0.857(95％CI:0.764～0.950)in the testing set.The nomogram construc-ted combined the clinical model and multimodal radiomics model had better predictive performance,with an AUC of 0.960(95％CI:0.933～0.987)in the training set and 0.890(95％CI:0.812'～0.968)in the testing set.DCA revealed that the nomogram provided a high clinical net benefit.Conclusion:The combination of multimodal radiomics and machine learning models can more accurately predict CLNM in patients with PTC,which has certain clinical application value.
维普
万方数据