基于多种机器学习算法的MRI影像组学模型预测直肠癌微卫星不稳定状态  

Prediction of microsatellite instability in rectal cancer using MRI radiomics models based on multiple machine learning algorithms

作　　者：李陆 卜文 孙巧玉 王伟 张玉文 姜海东 陈艾琪 沈俊杰 LI Lu;BU Wen;SUN Qiaoyu;WANG Wei;ZHANG Yuwen;JIANG Haidong;CHEN Aiqi;SHEN Junjie(Department of Radiology,The First Affiliated Hospital of Bengbu Medical University,Bengbu 233000,Anhui Province,China)

机构地区：[1]蚌埠医科大学第一附属医院放射科,安徽蚌埠233000

出　　处：《肿瘤影像学》2024年第6期577-585,共9页Oncoradiology

摘　　要：目的:探究联合临床危险因素、多参数磁共振成像(magnetic resonance imaging,MRI)影像组学特征构建的多种影像组学模型预测直肠癌微卫星不稳定(microsatellite instability,MSI)状态的价值。方法:纳入2020年12月—2023年11月蚌埠医科大学第一附属医院149例直肠癌患者,其中34例为MSI,115例为微卫星稳定(microsatellite stability,MSS)。基于MRI检查的多序列图像勾画感兴趣区,提取组学特征降维筛选出最佳特征,然后使用逻辑回归(logistic regression,LR)、随机森林(random forest,RF)、支持向量机(support vector machine,SVM)、K最近邻(k-nearest neighbor,KNN),朴素贝叶斯(naive Bayes tree,NBT)5种不同的机器学习算法将最优组学特征用于构建不同的影像组学模型,并绘制出受试者工作特征(receiver operating characteristic,ROC)曲线评估不同模型的诊断效能。结果:RF模型表现最稳定,且基于临床独立危险因素及影像组学构建的临床-影像组学联合模型列线图展现出对直肠癌MSI较高的诊断效能,训练组和验证组曲线下面积(area under curve,AUC)分别为0.923、0.914,评估直肠癌MSI效果最为显著。结论:结合不同机器学习算法,由临床危险因素和多参数MRI影像组学特征构建的临床-影像组学列线图可以有效地预测术前直肠癌MSI状态。Objective:To investigate the value of multiple radiomics models constructed by combining clinical risk factors and multiparametric magnetic resonance imaging(MRI)radiomics features to predict microsatellite instability(MSI)in rectal cancer.Methods:A total of 149 rectal cancer patients were included in the First Affiliated Hospital of Bengbu Medical University from December 2020 to November 2023,including 34 patients with MSI and 115 patients with microsatellite stability(MSS).Based on MRI examination of multiple sequence images,3D regions of interest were delineated,radiomics features were extracted and dimensionality was reduced to select the best features.Then,five different machine learning algorithms,including logistic regression(LR),random forest(RF),support vector machine(SVM),k-nearest neighbor(KNN),and naive Bayes tree(NBT)were used to construct different imaging radiomics models using the optimal radiomics features.And receiver operating characteristic(ROC)curves were drew to evaluate the diagnostic performance of different models.Results:The RF model showed the most stable performance,and the clinical imaging radiomics joint model nomogram based on clinical independent risk factors and imaging radiomics showed high diagnostic efficiency for MSI in rectal cancer.The area under curve(AUC)of the training group and the validation group were 0.923 and 0.914,respectively,indicating the most significant evaluation of MSI in rectal cancer.Conclusion:Combining different machine learning algorithms,a clinical imaging radiomics nomogram constructed from clinical risk factors and multi-parameter MRI radiomics features can effectively predict the unstable state of preoperative rectal cancer.

关 键 词：直肠癌 微卫星不稳定性 磁共振成像 影像组学 列线图 

分 类 号：R735.37[医药卫生—肿瘤] R445.2[医药卫生—临床医学]