基于机器学习的CT平扫影像组学鉴别CE1型肝囊型包虫病与肝囊肿  

Differential Diagnosis of Hepatic Cystic Echinococcosis(CE1)and Hepatic Syst Based on Radiomics Features of PlainCT and Machine Learning

作　　者：汪占金 曹云太 乔木 许晓磊 周瀛[2] 王海久[2] 杜凯豪 任悠悠 耿云平[4] 范建锋 王展[2] WANG Zhanjin;CAO Yuntai;QIAO Mu(School of Clinical Medicine,QinghaiUniversity,Xining,Qinghai,810000,P.R.China)

机构地区：[1]青海大学临床医学院,810001 [2]青海大学附属医院,810000 [3]郑州大学第三附属医院,451150 [4]南阳市中心医院,473000 [5]南阳市镇平县人民医院,473000

出　　处：《临床放射学杂志》2024年第9期1512-1518,共7页Journal of Clinical Radiology

基　　金：国家自然科学基金项目(编号:82160131);青海省科技厅项目(编号:2021-ZJ-963Q)。

摘　　要：目的开发CT平扫影像组学的机器学习分类器,鉴别CE1型肝囊型包虫病与肝囊肿。方法回顾性搜集接受手术治疗的48例CE1型肝囊型包虫病患者的CT平扫影像资料以及接受治疗的119例肝囊肿患者的CT平扫影像资料用于开发分类器。同时搜集果洛和玉树地区的24例CE1型肝囊型包虫病患者与28例肝囊肿患者的CT平扫影像资料进行外部验证。通过Python进行影像标准化和重采样,以减少因不同设备和参数引起的偏倚。采用LASSO回归筛选影像特征,并随机以7∶3的比例将其分为训练集和测试集。建立包括逻辑回归(LR)、决策树(DT)、随机森林(RF)、K近邻(KNN)、Catboost、XGBoost、LightGBM、Adaboost、多层感知器(MLP)、朴素贝叶斯(GNB)等10种机器学习算法的分类器。通过3折交叉网格搜索调整超参数,优化分类器性能,最终使用AUC值、准确率、F1值、敏感度、特异度评估在测试集和验证集上的分类器性能。结果LASSO回归筛选出14个影像特征,包括1个形态特征、10个纹理特征和3个一阶特征。在测试集中,Catboost分类器表现出最佳性能,AUC值为97.9%,准确率为96.1%。在外部验证中,Catboost分类器的表现最优,AUC值为0.932,准确率为90.4%。结论基于平扫CT影像组学区分肝囊型包虫(CE1)与肝囊肿是可行的,具有很大的潜力。Catboost分类器表现出卓越的性能,具有较强的泛化能力。Objective This study aims to analyze the imaging features of CE1-type hepatic cystic echinococcosis and liver cysts on CT plain scan images,with the goal of establishing a machine learning classifier for predictive purposes.Methods The study retrospectively gathered CT plain scan image data from 48 patients diagnosed with CE1-type hepatic cystic echinococcosis who underwent surgical treatment.Additionally,data were collected from 119 patients with liver cysts treated.These datasets were utilized for developing the classifier.Furthermore,CT plain scan image features from 24 CE1-type hepatic cystic echinococcosis patients and 28 liver cyst patients in the Guoluo Yushu area were collected for external validation.Image standardization and resampling were conducted using Python to mitigate bias arising from diverse devices and parameters.LASSO regression was employed for image feature selection,and the features were partitioned into training and testing sets in a 7∶3 ratio.A classifier incorporating ten machine learning algorithms logistic regression(LR),decision tree(DT),random forest(RF),k-nearest neighbors(KNN),Catboost,XGBoost,LightGBM,Adaboost,multilayer perceptron(MLP),and naive Bayes(GNB)was constructed.Hyperparameters were fine-tuned through a 3-fold cross-grid search to optimize classifier performance.Evaluation of the classifier's performance on the testing and validation sets involved metrics such as AUC,accuracy,F1 score,sensitivity,and specificity.Results LASSO regression identified 15 image features,encompassing 1 morphological feature,10 texture features,and 3 first-order features.Notably,10 features underwent wavelet filtering.In the testing set,the Catboost classifier demonstrated superior performance with an AUC of 0.979 and accuracy of 0.961.During external validation,the Catboost classifier showcased optimal results with an area under the ROC curve of 0.932 and accuracy of 0.904.Conclusion Conducting imaging omics research on cystic echinococcosis using CT plain scan is deemed feasible.The incorp

关 键 词：肝囊型包虫 肝囊肿 影像组学 机器学习 

分 类 号：R532.32[医药卫生—内科学] R575.8[医药卫生—临床医学] R816.5