基于磁共振成像机器学习算法和影像组学构建预测模型预测无淋巴结转移直肠癌淋巴血管侵犯状态的应用价值

Application value of prediction model based on magnetic resonance imaging machine learning algorithm and radiomics in predicting lymphovascular invasion status of rectal cancer without lymph node metastasis

目的基于磁共振成像(MRI)机器学习算法和影像组学构建预测模型,探讨模型预测无淋巴结转移直肠癌患者淋巴血管侵犯(LVI)状态的应用价值.方法采用回顾性队列研究方法.收集2016年2月至2024年1月甘肃省人民医院收治204例无淋巴结转移直肠癌患者的临床病理资料;男123例,女81例;年龄为(61±7)岁.204例患者采用电子计算器随机法按8∶2随机分为训练集163例和测试集41例.训练集用于构建预测模型,测试集用于验证预测模型效能.根据筛选的临床和(或)影像学特征分别构建临床预测模型、影像组学模型、联合预测模型.正态分布的计量资料以(x)±s表示.计数资料以绝对数表示,组间比较采用χ^(2)检验或Fisher确切概率法.等级资料比较采用非参数秩和检验.采用组间相关系数评估2位医师影像组学特征的一致性,相关系数>0.80认为一致性较好.单因素分析采用相应的统计学方法.多因素分析采用Logistic逐步回归模型.绘制受试者工作特征(ROC)曲线,以曲线下面积(AUC)及Delong检验、决策曲线和临床影响曲线评估模型的诊断效能及临床效用.结果(1)影响患者LVI状态的因素分析.204例无淋巴结转移直肠癌患者中,LVI阳性71例,LVI阴性133例.多因素分析结果显示:性别、血小板(PLT)计数和癌胚抗原(CEA)是影响训练集无淋巴结转移直肠癌患者LVI状态的独立因素[优势比=2.405,25.062,2.528,95%可信区间(CI)为1.093~5.291,2.748~228.604,1.181~5.410,P<0.05].(2)临床预测模型建立.纳入多因素分析结果性别、PLT计数和CEA构建临床预测模型.ROC曲线显示:临床预测模型训练集的AUC、准确度、灵敏度、特异度分别为0.721(95%CI为0.637~0.805)、0.675、0.632、0.698;测试集上述指标分别为0.795(95%CI为0.644~0.946)、0.805、1.000、0.429.Delong检验结果显示:训练集和测试集AUC比较,差异无统计学意义(Z=-0.836,P>0.05).(3)影像组学模型建立.提取204例患者851个影像组学特征.利用逻辑回归、支持向量机、高斯过程、逻辑回归-套索算法、线性判别分析、朴素贝叶斯和自动编码器7种机器学习算法进行预测模型构建.最终从最优的高斯过程机器学习算法中筛选出8个影像组学特征用于构建影像组学预测模型.ROC曲线显示:影像组学预测模型训练集的AUC、准确度、灵敏度、特异度分别为0.857(95%CI为0.800~0.914)、0.748、0.947、0.642;测试集上述指标分别为0.725(95%CI为0.571~0.878)、0.634、1.000、0.444.Delong检验结果显示:训练集和测试集AUC比较,差异无统计学意义(Z=1.578,P>0.05).(4)联合预测模型建立.联合多因素分析结果和影像组学特征构建联合预测模型.ROC曲线显示:联合预测模型训练集的AUC、准确度、灵敏度、特异度分别为0.885(95%CI为0.832~0.938)、0.791、0.912、0.726;测试集上述指标分别为0.857(95%CI为0.731~0.984)、0.854、0.714、0.926.Delong检验结果显示:训练集和测试集AUC比较,差异无统计学意义(Z=0.395,P>0.05).(5)3种预测模型效能比较.Hosmer-Lemeshow拟合优度检验结果显示:临床预测模型、影像组学预测模型、联合预测模型的拟合优度均较好(χ^(2)=1.464,12.763,10.828,P>0.05).Delong检验结果显示:临床预测模型AUC分别与联合预测模型和影像组学预测模型比较,差异均无统计学意义(Z=1.146,0.658,P>0.05);联合预测模型与影像组学模型比较,差异有统计学意义(Z=2.001,P<0.05).校准曲线显示:联合预测模型的预测能力良好.决策曲线和临床影响曲线显示:联合预测模型对无淋巴结转移直肠癌患者LVI状态的评估能力优于临床预测模型和影像组学模型.结论纳入性别、PLT计数和CEA构建临床预测模型;筛选出8个影像组学特征构建影像组学预测模型;结合前两者共同构建联合预测模型.3种模型均可预测无淋巴结转移直肠癌患者LVI状态,其中联合预测模型效能更优.

Objective To construct an prediction model based on magnetic resonance imaging(MRI)machine learning algorithm and radiomics and investigate its application value in predicting lymphovascular invasion(LVI)status of rectal cancer without lymph node metastasis.Methods The retrospective cohort study was conducted.The clinicopathological data of 204 rectal cancer patients without lymph node metastasis who were admitted to Gansu Provincial Hospital from February 2016 to January 2024 were collected.There were 123 males and 81 females,aged(61±7)years.All 204 patients were randomly divided into the training dataset of 163 cases and the testing dataset of 41 cases by a ratio of 8:2 using the electronic computer randomization method.The training dataset was used to construct the prediction model,and the testing dataset was used to validate the prediction model.The clinical prediction model,radiomics model and joint prediction model were constructed based on the selected clinical and/or imaging features.Measurement data with normal distribution were represented as Mean+SD.Count data were described as absolute numbers,and the chi-square test or Fisher exact probability were used for comparison between the groups.Comparison of ordinal data was conducted using the nonparameter rank sum test.The inter-class correlation coefficient(ICC)was used to evaluate the consistency of the radiomics features of the two doctors,and ICC>0.80 was good consistency.Univariate analysis was conducted by corres-ponding statistic methods.Multivariate analysis was conducted by Logistic stepwise regression model.The receiver operating characteristic(RoC)curve was drawn,and the area under the curve(AUC),Delong test,decision curve and clinical impact curve were used to evaluate the diagnostic efficiency and clinical utility of the model.Result(1)Analysis of factors affecting LVI status of patients.Of the 204 rectal cancer patients without lymph node metastasis,there were 71 cases with positive of LVI and 133 cases with negative of LVI.Results of multivariate analysis showed that gender,platelet(PLT)count and carcinoembryonic antigen(CEA)were independent factors affecting LVI status of rectal cancer without lymph node metastasis in training dataset[odds ratio=2.405,25.062,2.528,95%confidence interval(C)as 1.093-5.291,2.748-228.604,1.181-5.410,P<0.05].(2)Construction of clinical prediction model.The clinical prediction model was conducted based on the results of multivariate analysis including gender,PLT count and CEA.Results of ROC curve showed that the AUC,accuracy,sensitivity and specificity of clinical prediction model were 0.721(95%Cl as 0.637-0.805),0.675,0.632 and 0.698 for the training dataset,and 0.795(95%Cl as 0.644-0.946),0.805,1.000 and 0.429 for the testing dataset.Results of Delong test showed that there was no significant difference in the AUC of clinical prediction model between the training dataset and the testing dataset(Z=-0.836,P>0.05).(3)Construction of radiomics model.A total of 851 radiomics features were extracted from 204 patients,and seven machine learning algorithms,including logistic regression,support vector machine,Gaussian process,logistic regression-lasso algorithm,linear discriminant analysis,naive Bayes and automatic encoder,were used to construct the prediction model.Eight radiomics features were finally selected from the optimal Gaussian process learning algorithm to construct a radiomics prediction model.Results of RoC curve showed that the AUC,accuracy,sensitivity and specificity of radiomics prediction model were 0.857(95%CI as 0.800-0.914),0.748,0.947 and 0.642 for the training dataset,and 0.725(95%CI as 0.571-0.878),0.634,1.000 and 0.444 for the testing dataset.Results of Delong test showed that there was no significant difference in the AUC of radiomics prediction model between the training dataset and the testing dataset(Z=1.578,P>0.05).(4)Construction of joint prediction model.The joint prediction model was constructed based on the results of multivariate analysis and the radiomics features.Results of ROC curve showed that the AUC,accuracy,sensitivity and specificity of radiomics prediction model were 0.885(95%CI as 0.832-0.938),0.791,0.912 and 0.726 for the training dataset,and 0.857(95%CI as 0.731-0.984),0.854,0.714 and 0.926 for the testing dataset.Results of Delong test showed that there was no significant difference in the AuC of joint prediction model between the training dataset and the testing dataset(Z=0.395,P>0.05).(5)Performance comparison of three prediction models.Results of the Hosmer-Lemeshow goodness-of-fit test showed that all of the clinical prediction model,radiomics prodiction model and joint prediction model having good fitting degree(χ^(2)=1.464,12.763,10.828,P>0.05).Results of Delong test showed that there was no significant difference in the AUC between the clinical prediction model and the joint prediction model or the radiomics model(Z=1.146,0.658,P>0.05),and there was a significant difference in the AUC between the joint prediction model and the radiomics model(Z=2.001,P<0.05).Results of calibration curve showed a good performance in the joint prediction model.Results of decision curve and clinical impact curve showed that the performance of joint prediction model in predicting LVI status of rectal cancer without lymph node metastasis was superior to the clinical prediction model and the radiomics model.Conclusions The clinical prediction model is constructed based on gender,PLT count and CEA.The radiomics predictive model is constructed based on 8 selected radiomics features.The joint prediction model is constructed based on the clinical prediction model and the radiomics predictive model.All of the three models can predict the LVI status of rectal cancer without lymph node metastasis,and the joint prediction model has a superior predictive performance.