基于CT平扫影像组学特征机器学习算法模型对高级别Bosniak分类囊性肾占位性病变的诊断效能

Value of machine learning algorithm model based on CT plain scan radiomics in diagnosis for cystic renal mass of high-grade Bosniak classification

目的:比较基于CT平扫影像组学特征的机器学习算法模型与放射科医师对囊性肾占位性病变(CRM)的诊断效能。方法:招募来自广东省中医院5家分院经病理诊断的CRM患者207例,其中4家162例(内部测试集)按7∶3比例随机分为训练集与验证集,训练集提取CT平扫影像组学特征用于机器学习算法建模,由2位放射科医师进行全肿瘤分割。使用单变量分析、最小绝对收缩和选择算子(LASSO)算法及双向消除法筛选ICC>0.75的特征,以构建朴素贝叶斯、梯度提升算法(GBM)及主成分分析(PCA)的监督与非监督影像组学机器学习算法模型,进行验证集验证,在珠海分院45例(外部测试集)中完成外部测试。使用ROC曲线评估机器学习算法模型诊断效能,并与放射科医师诊断结果进行比较。结果:207例中,良性92例,恶性115例。放射科医师诊断的敏感度、特异度、准确率分别为84.2%、91.1%和85.5%,AUC为0.87。机器学习算法模型在训练集和验证集中诊断效能相近[训练集敏感度(95.3%~100.0%)、特异度(93.4%~100.0%)、准确率(94.4%~100.0%)、AUC(0.97~1.00);验证集敏感度(83.3%~89.9%)、特异度(87.0%~95.7%)、准确率(85.4%~90.2%)、AUC(0.85~0.92)];在外部测试集中,PCA和朴素贝叶斯的敏感度、特异度、准确率和AUC均高于GBM。放射科医师的诊断效能与PCA及朴素贝叶斯模型相近。结论:CT平扫影像组学的多种机器学习算法模型对CRM具有较好的诊断效能,可作为CRM患者的潜在筛查方法。

Objective:To compare the diagnostic performance of machine learning(ML)algorithm model based on CT plain scan radiomics and radiologists in the evaluation of cystic renal mass(CRM).Methods:207 CRM patients from five branches of our hospital were enrolled,and 162 CRM patients from four branches were randomly divided into a training and validation set in a ratio of 7∶3.Radiomics features were extracted from CT plain scan for ML modeling,and the entire tumor was segmented by two radiologists.Radiomics features with an ICC>0.75 were selected by using univariate analysis,least absolute shrinkage and selection operator(LASSO)and bidirectional elimination to build supervised and unsupervised ML radiomics models with Naive Bayes,Gradient Boosting Machines(GBM)and Principal Component Analysis(PCA).These models were validated with the validation set and then externally tested in 45 CRM patients from the fifth branch.The ML models were evaluated using ROC curves,and compared with radiologists'diagnosis.Results:In 207 patients,92 cases were benign CRM,and 115 cases malignant.The sensitivity,specificity and accuracy of the radiologists'diagnosis were 84.2%,91.1%and 85.5%,respectively,and AUC was 0.87.The ML algorithm model showed comparable performance in the training and validation sets,the sensitivity,specificity,accuracy and AUC of the training set were 95.3%~100.0%,93.4%~100.0%,94.4%~100.0%and 0.97~1.00,and those of the validation set were 83.3%~89.9%,87.0%~95.7%,85.4%~90.2%and 0.85~0.92,respectively.In the external test set,PCA and Naive Bayes showed better sensitivity,specificity,accuracy and AUC compared to GBM.Conclusion:Multiple ML algorithms models based on CT plain scan radiomics have good diagnostic efficacy in evaluating CRM,offering a potential screening method for CRM patients.