超声影像组学在肝包虫病分型中的应用研究

A pplication of ultrasound radiomics in the classification of hepatic echinococcosis

目的 基于超声影像组学构建机器学习模型,探究其在鉴别囊型肝包虫病与泡型肝包虫病中的应用价值。方法 选取于我院及四川省甘孜藏族自治州石渠县经病理结果(金标准)或“四川省包虫病专家组”依据临床结果及专家共识(银标准)确诊的肝包虫病患者4976例,共纳入23 452张超声图像,其中囊型肝包虫病图像8557张,泡型肝包虫病图像14 895张。按病灶类型以8∶2比例将超声图像随机分为训练集18 762张与独立测试集4690张。使用Pyradiomics(3.1.0)提取超声图像的影像组学特征,对训练集及独立测试集超声图像均使用相同的提取器;使用极小化极大、标准分数及均值方法对影像组学特征进行特征缩放,使用主成分分析和Pearson相关系数进行特征降维,使用方差分析、递归特征消除、相关特征法及克鲁斯卡尔-沃利斯法筛选最佳影像组学特征。基于支持向量机(SVM)、自编码器(AE)、线性判别分析(LDA)、随机森林(RF)、逻辑回归(LR)、自适应增强(AB)、决策树(DT)、朴素贝叶斯(NB)8种分类器构建鉴别肝包虫病分型的最佳机器学习模型。训练集中,以十折交叉验证策略训练模型。绘制受试者工作特征(ROC)曲线分析不同分类器中最佳机器学习模型在训练集及独立测试集中鉴别肝包虫病分型的诊断效能。结果 从每张超声图像中共提取1130个影像组学特征,经特征选择动态筛选出1~40个最佳影像组学特征,并以此建立最佳机器学习模型。ROC曲线分析显示,RF模型在训练集及独立测试集中鉴别肝包虫病分型的曲线下面积分别为0.82、0.86,均高于SVM、AE、LDA、LR、AB、DT、NB模型,差异均有统计学意义(均P<0.05)。结论 基于超声影像组学的RF模型对鉴别肝包虫病分型的诊断效能最佳,有助于肝包虫病精准超声诊断。

Objective To construct a machine learning model based on ultrasound radiomics,and to explore the application value in the differentiating cystic echinococcosis and alveolar echinococcosis.Methods A total of 4976 patients diagnosed with hepatic echinococcosis in our hospital and Shiqu county,Ganzi tibetan autonomous prefecture,Sichuan province according to pathological results(gold standard)or diagnosed by the“Sichuan Province Echinococcosis Expert Group”based on clinical outcomes and expert consensus(silver standard)were selected,including a total of 23452 ultrasound images,with 8557 images of cystic echinococcosis and 14895 images of alveolar echinococcosis.The ultrasound images were randomly divided into training set(18762 images)and independent test set(4690 images)in a 8∶2 ratio according to the lesion type.Pyradiomics(3.1.0)was used to extract radiomic features from ultrasonographic images,and the same extractor was applied for the ultrasonic images of the training set and the independent test set.MinMax,Z-score and Mean methods were used for feature scaling of radiomic features.Principal component analysis and Pearson correlation coefficient were used for feature dimensionality reduction,and analysis of variance,recursive feature elimination,relevant features,as well as the Kruskal-Wallis methods were used to screen the best image radiomic features.The machine learning models were constructed based on 8 classifiers,including support vector machine(SVM),auto-encoder(AE),linear discriminant analysis(LDA),random forest(RF),Logistic regression(LR),adaptive boosting(AB),decision tree(DT)and naive Bayes(NB).In the training set,a ten-fold cross-validation strategy was employed to train the model.Receiver operating characteristic(ROC)curve was drawn to analyze the diagnostic performance of the best machine learning model in different classifiers in the training set and the independent test set in differentiating the classification of hepatic echinococcosis.Results A total of 1130 radiomic features were extracted from each ultrasonic image,and 1~40 optimal radiomic features were dynamically selected by feature selection to establish the optimal machine learning models.ROC curve analysis showed that the area under the curve of RF model for the classification of hepatic echinococcosis in the training and independent test sets were 0.82 and 0.86,respectively,which were higher than those of the SVM,AE,LDA,LR,AB,DT and NB models,and the differences were statistically significant(all P<0.05).Conclusion The RF model based on ultrasound radiomics demonstrates the optimal diagnostic efficacy in differentiating the classification of hepatic echinococcosis,which is helpful for the precise ultrasound diagnosis of the disease.