基于MRI影像组学的机器学习模型预测前列腺癌骨转移的价值

The value of machine learning model for predicting prostate cancer bone metastases based on MRI radiomics

目的本研究旨在建立并验证基于MRI的机器学习模型用于预测前列腺癌(prostate cancer,PCa)骨转移的价值。材料与方法回顾性分析2018年1月至2022年1月来自山东第一医科大学第一附属医院和山东第一医科大学附属省立医院的150名经病理证实的PCa患者的临床和MRI资料,按照7∶3的比例随机分为训练集(105例)和测试集(45例),分别在每个患者的表观扩散系数(apparent diffusion coefficient,ADC)和T2脂肪抑制序列(fat saturated T2 weighted imaging,FS-T2WI)图像手动勾画肿瘤的感兴趣区并提取影像组学特征。使用组内相关系数、特征重要性及最小冗余最大相关性方法进行降维和特征筛选,筛选出的最佳特征用广义线性模型(generalized linear model,GLM)、随机森林(random forest,RF)、支持向量机(support vector machine,SVM)和极致梯度提升(extreme gradient boosting,XGB)方法建立影像组学模型。采用受试者工作特征曲线下面积(area under the curve,AUC)、校准曲线和决策曲线分析去评估四种模型的预测性能,然后通过德隆检验比较模型之间的差异。结果获得了17个最佳特征构建机器学习模型,在训练集中,GLM、XGB、SVM和RF模型的平均AUC分别为0.714、0.845、0.768和0.858;在测试集中,对应的AUC分别为0.796、0.729、0.755和0.765。德隆检验和校准曲线表示四种模型之间没有显著差异。决策曲线分析显示四种模型的临床应用价值相近。结论基于MRI影像组学特征的GLM、XGB、SVM和RF模型能够作为一种有前景的工具预测PCa患者骨转移,为临床诊疗提供潜在的有效信息。

Objective:To develop and validate MRI-based machine learning models for predicting bone metastases in patients with prostate cancer(PCa).Materials and Methods:The clinical and MRI data of 150 patients with pathologically confirmed PCa in the First Affiliated Hospital and Affiliated Provincial Hospital of Shandong First Medical University were retrospectively obtained from January 2018 to January 2022.According to the ratio of 7∶3,the samples were randomly divided into training set(n=105)and testing set(n=45).The apparent diffusion coefficient(ADC)and fat saturated T2 weighted imaging(FS-T2WI)of each patient were manually outlined for the tumor’s region of interest and extracted for imaging histological features,respectively.Dimension reduction and feature selection were performed using intra-class correlation coefficients(ICC),feature importance and minimal-redundancy-maximal-relevance(mRMR).The filtered features were used to establish radiomics models using generalized linear model(GLM),random forest(RF),support vector machine(SVM)and extreme gradient boosting(XGB)methods.The models were evaluated using the following metrics:area under the curve(AUC)of receiver operating characteristic(ROC),calibration curve,decision curve analysis and Delong test.Results:Seventeen features were selected and models were constructed using GLM,XGB,SVM and RF.In the training set,the mean AUC were 0.714,0.845,0.768 and 0.858,respectively.In the testing set,the AUC were 0.796,0.729,0.755 and 0.765,respectively.Calibration curve and Delong test indicated no significant differences between the four models.Decision curve analysis showed that the four models had similar clinical applications.Conclusions:The MRI-based radiomics features allowed GLM,SVM,XGB and RF classifiers to be used as a promising tool for predicting bone metastases in PCa patients,providing potentially valid information for clinical management.