摘要
目的:探讨基于多参数磁共振成像(multi-participant magnetic resonance imaging,mpMRI)的影像组学模型对前列腺癌Gleason分级的应用价值。方法:回顾并分析2020年11月—2023年8月在马鞍山市人民医院行前列腺MRI检查且经穿刺活检或术后病理学检查证实为前列腺癌的患者资料。提取mpMRI图像数据,包括T2加权成像(T2-weighted imaging,T2WI)、小视野弥散加权成像(zoomed imaging technique with parallel transmission diffusion-weighted imaging,ZOOMit DWI)、表观弥散系数(apparent diffusion coefficient,ADC)。采用Spearman相关系数初步筛选组学特征,采用最小绝对收缩和选择算子(least absolute shrinkage and selection operator,LASSO)算法及10折交叉验证进一步筛选,采用logistic回归构建模型,使用受试者工作特征(receiver operating characteristic,ROC)曲线判断模型的诊断效能,使用DeLong检验比较模型间曲线下面积(area under curve,AUC)。结果:共纳入176例患者,包括低级别组72例(Gleason评分≤3+4),高级别组104例(Gleason评分≥4+3),按7∶3随机分成训练集(n=141)和测试集(n=35)。应用多种分类器对多参数模型进行构建,结果显示支持向量机(support vector machine,SVM)在测试集中AUC为0.891,训练集中AUC为0.905。轻量级梯度提升机(light gradient boosting machine,LightGBM)在训练集中AUC最高,为0.931;但其在测试集中表现欠佳,AUC为0.808。多层感知机(multilayer perceptron,MLP)在测试集和训练集中AUC可分别达到0.883、0.855,整体弱于SVM;可见LightGBM和MLP模型稳定性相较于SVM来说略差。另外K紧邻(k-nearest neighbor,KNN)、极度随机树(extra trees,ET)、随机森林(random forest,RF)、极度梯度提升(extreme gradient boosting,XGBoost)这4种方法的整体效能也都不如SVM,且部分存在过拟合。综合而言,在前列腺癌Gleason分级方面,SVM模型无论测试集还是训练集AUC均较高,其稳定性以及模型分级能力更好。结论:基于mpMRI构建多模态影像组学模型在前列腺癌Gleason分级中有较大的临床应用价值,其中以SVM模型为最佳。
Objective:To explore the application value of radiomics model based on multi-parameter magnetic resonance imaging(mpMRI)in Gleason grading of prostate cancer.Methods:The data of patients who underwent prostate mpMRI examination with surgical or pathological puncture results confirming prostate cancer at Ma’anshan People’s Hospital from November 2020 to August 2023 were retrospectively analyzed.MpMRI data were extracted,including T2-weighted imaging(T2WI),zoomed imaging technique with parallel transmission diffusion-weighted imaging(ZOOMit DWI)and apparent diffusion coefficients(ADC).Spearman’s correlation coefficient was used to preliminarily screen the histological features,the least absolute shrinkage and selection operator(LASSO)algorithm and ten-fold cross-validation were used to further screen,logistic regression was used to construct the model,and the receiver operating characteristic(ROC)curve was used to judge the results.And the area under the ROC curve(AUC)was compared between models using the DeLong test.Results:A total of 176 patients were included,including 72 patients in the low-grade group(Gleason score≤3+4)and 104 patients in the high-grade group(Gleason score≥4+3),who were randomly divided into training group(n=141)and test group(n=35)according to 7∶3.A variety of classifiers were used to construct the multi-parameter model,and the results showed that the AUC of support vector machine(SVM)in the test set was 0.891,and the AUC in the training set was 0.905.Light gradient boosting machine(LightGBM)had the highest AUC of 0.931 in the training set,but it performed poorly in the test set with an AUC of 0.808.The AUCs of multilayer perceptron(MLP)in the test set and the training set were 0.883 and 0.855,respectively,which were weaker than that of SVM,which showed that the stability of LightGBM and MLP models were slightly worse than that of SVM.In addition,the overall performance of the four methods[k-nearest neighbor(KNN),extra trees(ET),random forest(RF),extreme gradient boosting(XGBoost)]were not as good as SVM,and some of them are overfitted.In general,in terms of Gleason grading of prostate cancer,the SVM model had a higher AUC in both the test set and the training set,and its stability and model classification ability were better.Conclusion:Constructing a multimodal imaging histology model based on mpMRI has significant clinical application value in Gleason grading of prostate cancer,of which the SVM model is the best.
作者
杨馨
杨宏楷
戚轩
翟承凤
何永胜
YANG Xin;YANG Hongkai;QI Xuan;ZHAI Chengfeng;HE Yongsheng(The Fifth Clinical Medical College of Anhui Medical University,Ma’anshan Clinical College of Anhui Medical University,Ma’anshan 243000,Anhui Province,China;Department of Imaging,Ma’anshan People’s Hospital,Ma’anshan 243000,Anhui Province,China)
出处
《肿瘤影像学》
2024年第5期536-544,共9页
Oncoradiology
基金
安徽省重点研究与开发计划(2022e07020065)。
