摘要
目的 探讨基于MRI影像组学特征的机器学习分类器诊断半月板不同损伤级别的价值.方法 回顾性分析368例患者膝关节MRI图像.联合矢状位及冠状位质子密度加权脂肪抑制图像的影像组学特征,经Select-KBest、最小冗余最大相关和最小绝对收缩和选择算子回归进行特征筛选和降维,选择最优特征,采用多种机器学习方法构建半月板损伤四分类诊断模型,应用受试者工作特征曲线评估各模型诊断效能.结果 最终获得18个最优特征,支持向量机、逻辑回归、高斯过程、随机森林、二次判别分析和Bagging决策树模型的AUC值最佳,分别为0.876、0.871、0.870、0.869、0.868和0.868.随机森林、逻辑回归、Bagging决策树和随机森林分别在诊断正常半月板、1、2、3级半月板损伤中AUC值最高(0.948、0.833、0.805、0.902).结论 基于MRI影像组学特征的机器学习模型进行半月板损伤分级的可行性较好,具有良好的诊断效能.
Abstract
Objective To explore the value of machine learning models based on MRI radiomics features in predicting the degree of meniscus damage.Methods Knee MRI images of 368 patients were retrospectively analyzed.Combined with sagittal and coronal proton density-weighted fat suppression images,the SelectKBest,Minimum redundancy maximum relevance and least absolute shrinkage and selection operator were used to select and reduce the dimension of the radiomics features.Then,based on the optimal features,a variety of machine learning methods are used to build a four-category pre-diction model,and its performance was evaluated by the receiver operating characteristic(ROC)curve.Results Finally,18 optimal features are obtained.The Macro AUC values of support vector machine,logistic regression,Gaussian process,random forest,quadratic discriminant analysis and Bagging decision tree model were the best,which were 0.876,0.871,0.870,0.869,0.868 and 0.868,respectively.The AUC values of random forest,logistic regression,Bagging decision tree and random forest were the highest(0.948,0.833,0.805,0.902)in the diagnosis of normal meniscus injury,grade 1,grade 2,grade 3 meniscus damage,respectively.Conclusion The machine learning model based on MRI radiomics fea-tures for meniscus damage is feasible and has good diagnostic efficiency.
基金项目
天津市医学重点学科(专科)建设项目(TJYXZDXK-026A)
NETL
NSTL
万方数据