目的 探索基于MRI中T2WI序列影像组学中不同机器学习模型在宫颈鳞癌分化程度的预测效能.方法 回顾性分析2017年1月至2023年6月的119例宫颈鳞癌患者的临床资料和T2WI影像.两名放射科医生利用ITK-SNAP软件手动勾画病灶的三维感兴趣区域,使用Python编程语言提取1316个影像组学特征.采用t检验、LASSO回归筛选最优的影像组学特征,然后建立支持向量机(Support Vector Machine,SVM)、逻辑回归(Logistic Regression,LR)、K近邻(K Nearest Neighbor,KNN)和随机森林(Random Forest,RF)共4种机器学习模型.采用受试者工作特征曲线下面积(Area Under the Curve,AUC)、分类准确度、敏感度和特异性来评估4种模型预测宫颈鳞癌分化程度的效能.结果 LR模型区分低分化与中分化、低分化与高分化、中分化与高分化的AUC分别为0.791、0.884、0.793;SVM模型的AUC分别为0.730、0.832、0.815;KNN模型的AUC分别为0.773、0.777、0.677;RF模型的AUC分别为0.824、0.763、0.783.结论 基于磁共振T2WI序列影像组学特征的4种机器学习模型,可作为术前预测宫颈鳞癌分化程度的辅助手段,但对不同分化程度的预测效能不同.
Predictive Efficacy of Four Machine Learning Models Based on T2WI MRI-Based for Differentiation Grade of Cervical Squamous Cell Carcinoma
Objective To explore the predictive efficacy of different machine learning models based on T2WI sequence MRI-based radiomics in the differentiation grade of cervical squamous cell carcinoma.Methods Clinical data and T2WI images of 119 patients with cervical squamous cell carcinoma from January 2017 to June 2023 were retrospectively analyzed.Two radiologists manually delineated three-dimensional region of interest of lesions using ITK-SNAP software,and extracted 1316 radiomics features using Python programming language.The optimal radiomics features were selected by t-test and LASSO regression,and four machine learning models of support vector machine(SVM),logistic regression(LR),K nearest neighbor(KNN)and random forest(RF)were established.The area under curve(AUC)of the receiver operating characteristic,classification accuracy,sensitivity and specificity were used to evaluate the performance of the four models in predicting the differentiation grade of cervical squamous cell carcinoma.Results The AUC of LR model was 0.791,0.884 and 0.793,the AUC of SVM model was 0.730,0.832 and 0.815,the AUC of KNN model was 0.773,0.777 and 0.677,and the AUC of RF model was 0.824,0.763 and 0.783,respectively in differentiating low differentiation from medium differentiation,low differentiation from high differentiation,medium differentiation from high differentiation.Conclusion Four machine learning models based on T2WI sequence MRI-based radiomics,can be used as an adjunct to preoperative prediction of the grade of differentiation of cervical squamous cell carcinoma,but their predictive efficacy varies for different degrees of differentiation.
国家科技期刊平台
NSTL
万方数据
