机器学习多种算法构建MRI影像组学模型预测睾丸良恶性病变的对比分析  

Comparative Analysis of Multiple Algorithms for Machine Learning to Construct MRI Imaging Histological Models to Predict Benign and Malignant Testicular Lesions

作　　者：简远熙 柳瑞[1] 鲁茸迪几 扎西七林 拉茸七林 尼玛卓玛 吴自强 杨素萍 JIAN Yuanxi;LIU Rui;Lurongdiji(Department of Radiology,The Second Affiliated Hospital of Kunming Medical University,Kunming,Yunnan Province 650101,P.R.China)

机构地区：[1]昆明医科大学第二附属医院放射科,650101 [2]香格里拉市人民医院影像科,674400

出　　处：《临床放射学杂志》2025年第4期689-694,共6页Journal of Clinical Radiology

摘　　要：目的探讨基于多参数MRI影像组学联合多种机器学习算法模型对睾丸良恶性病变的鉴别诊断价值。方法回顾性搜集148例经病理证实为睾丸良恶性病变患者的首次MRI及临床、实验室检查资料;其中良性60例,恶性88例;按照7∶3比例随机将其分为训练集(n=103)和测试集(n=45)。采用3D-Slicer软件逐层勾画病灶的感兴趣区(ROI),并提取组学特征。采用最大相关最小冗余(mRMR)、最小绝对收缩和选择算子算法(LASSO)筛选最优组学特征,并分别采用6种机器学习算法构建模型。应用受试者工作特征曲线(ROC)的曲线下面积(AUC)、敏感度、特异度、准确率评估模型的预测效能。结果对比增强(CE)-T_(1)WI+扩散加权成像(DWI)+T_(2)WI中提取的组学特征联合逻辑回归(LR)机器学习算法所构建的模型的AUC值在训练集和测试集分别为0.927和0.982,且敏感度、特异度和准确率为所有模型中最高。结论基于CE-T_(1)WI+DWI+T_(2)WI组学特征并结合LR机器学习算法的模型在睾丸良恶性病变的鉴别诊断中具有最佳的预测效能。Objective Exploring the identification of benign and malignant testicular lesions based on multiparametric MRI image histology combined with multiple machine learning algorithm models.Methods A retrospective collection of 148 cases of initial MRI and clinical and laboratory examinations of pathologically confirmed benign and malignant testicular lesions in our institution was performed.Of these,60 were benign cases and 88 were malignant cases.They were randomly divided into a training set(n=103)and a test set(n=45)according to a 7∶3 ratio.3D-Slicer software was used to outline the region of interest of the lesion layer by layer and extract histological features.Maximum correlation minimum redundancy,minimum absolute contraction and selection methods are used to screen the optimal features,and six machine learning algorithms were used to construct the model respectively.The predictive efficacy of the model was assessed by applying the area under the curve(AUC),sensitivity,specificity,and accuracy of receiver operating characteristic(ROC).Results The AUC values of the model constructed by the histological features extracted from CE-T_(1)WI+DWI+T_(2)WI in conjunction with the LR machine learning algorithm were 0.927 and 0.982 in the training and test sets,respectively,and had the highest sensitivity,specificity,and accuracy of all the models.Conclusion A model based on CE-T_(1)WI+DWI+T_(2)WI histological features and combined with LR machine learning algorithm has the best predictive efficacy in the differential diagnosis of benign and malignant testicular lesions.

关 键 词：睾丸病变 良性 恶性 磁共振成像 影像组学 

分 类 号：R73[医药卫生—肿瘤]