增强CT影像组学对结直肠癌微卫星不稳定状态的预测价值分析  

Analysis of the Predictive Value of Enhanced CT Imaging in Microsatellite Instability in Colorectal Cancer

作　　者：丁曙 徐圆 刘亚伟 滕越 陈望 孙关 DING Shu;XU Yuan;LIU Yawei(Department of Radiology,Yancheng No.1 People's Hospital,Affiliated Hospital of Medical School,Nanjing University(The First People’s Hospital of Yancheng),Yancheng,Jiangsu Province 224000,P.R.China)

机构地区：[1]南京大学医学院附属盐城第一医院(盐城市第一人民医院)影像科,盐城224000 [2]南京大学医学院附属盐城第一医院(盐城市第一人民医院)神经外科,盐城224000

出　　处：《临床放射学杂志》2025年第3期479-483,共5页Journal of Clinical Radiology

摘　　要：目的基于结直肠癌增强CT静脉期图像提取高通量影像组学参数,构建不同机器学习模型,探讨其对结直肠癌微卫星不稳定状态的预测价值。方法回顾性搜集277例结直肠癌患者的基本临床资料,其中,微卫星不稳定(MSI)组168例,微卫星稳定(MSS)组109例。采用单因素分析其统计学差异。应用Mazda软件对肿瘤最大层面进行感兴趣区的勾画并提取影像组学参数。将微卫星不稳定组及微卫星稳定组患者按7∶3的比例随机分为训练集及验证集(训练集MSI组118例及MSS组50例;验证集MSI组76例及MSS组33例)。结合五折交叉验证,绘制受试者工作特征曲线,评价支持向量机、线性判别分析、随机森林、逻辑斯回归、贝叶斯及决策树6类机器学习算法的诊断效能。结果患者的年龄、性别、吸烟史、肿瘤家族史、饮酒史、CEA、CA199及CA125两组间比较差异均不具有显著统计学意义(均P>0.05)。不同机器学习模型中,基于Relief算法及提取14个特征参数的贝叶斯模型诊断效能最好,敏感度、特异度、阳性预测值、阴性预测值、训练集平均曲线下面积及验证集平均曲线下面积分别为0.763、0.695、0.617、0.820、0.795及0.772。结论增强CT影像组学及贝叶斯模型在结直肠癌微卫星不稳定状态的预测中具有较高的诊断价值,为临床治疗决策提供了一定的参考价值。Objective To explore the predictive value of high-throughput radiomics parameters extracted from the venous phase of enhanced CT images of colorectal cancer using different machine learning models.Methods The basic clinical data of 277 patients with colorectal cancer were retrospectively collected,including 168 cases in the microsatellite instability(MSI)group and 109 cases in the microsatellite stable(MSS)group.Univariate analysis was used to identify statistical differences.The maximum tumor section was delineated and radiomics parameters were extracted using Mazda software.Patients in the MSI and MSS groups were randomly divided into training and validation sets in a 7:3 ratio(118 MSI and 50 MSS in the training set;76 MSI and 33 MSS in the validation set).Combined with five-fold cross-validation,receiver operating characteristic(ROC)curves were drawn to evaluate the diagnostic efficacy of six machine learning algorithms:support vector machine,linear discriminant analysis,random forest,logistic regression,Bayesian,and decision tree.Results There were no significant differences between the two groups in terms of patient age,gender,smoking history,family history of cancer,drinking history,CEA,CA199,and CA125 levels(all P values>0.05).Among the different machine learning models,the Bayesian model based on the Relief algorithm and the extraction of 14 feature parameters had the best diagnostic efficacy,with sensitivity of 0.763,specificity of 0.695,positive predictive value of 0.617,negative predictive value of 0.820,average AUC value in the training set of 0.795,and average AUC value in the validation set of 0.772.Conclusion Enhanced CT radiomics and Bayesian models have high diagnostic value in predicting microsatellite instability in colorectal cancer,providing a reference for clinical treatment decisions.

关 键 词：结直肠癌 影像组学 微卫星不稳定性 贝叶斯算法 电子计算机断层扫 

分 类 号：R735.34[医药卫生—肿瘤] R730.44[医药卫生—临床医学]