基于机器学习的影像组学模型对肌层浸润性膀胱癌的诊断价值  

Diagnostic value of machine learning-based MRI radiomics model in predicting muscle-invasive bladder cancer

作　　者：李陇超 常鸿志 杨艳蓉 折霞[1] 汤敏[1] 雷晓燕[1] 张小玲[1] 张沥[1] LI Long-chao;CHANG Hong-zhi;YANG Yan-rong;ZHE Xia;TANG Min;LEI Xiao-yan;ZHANG Xiao-ling;ZHANG Li(Department of MR,Shaanxi Provincial People’s Hospital,Shaanxi 710068,China)

机构地区：[1]陕西省人民医院磁共振室,陕西西安710068

出　　处：《影像诊断与介入放射学》2021年第3期212-216,共5页Diagnostic Imaging & Interventional Radiology

基　　金：陕西省重点研发计划(2018SF-169)。

摘　　要：目的建立基于Logistic回归分析法、R-Tree两种机器学习算法的影像组学模型,探讨各模型对肌层浸润性膀胱癌的诊断效能。方法回顾性分析经病理证实的膀胱癌132例,其中肌层浸润性膀胱癌51例,非肌层浸润性膀胱癌81例;所有患者术前均接受MRI检查,包括T2WI及DWI序列。两位影像科医生应用ITK-SNAP软件在T2WI和ADC图上手动勾画病灶的三维兴趣区(VOI)。应用AK软件对获得的VOI进行特征提取、降维、筛选,并基于不同机器学习算法构建三组预测模型(ADC、T2WI、T2WI+ADC)。采用受试者工作特征(ROC)曲线评估不同模型的诊断效能,获得曲线下面积(AUC)、符合率、特异度、敏感度。结果应用Logistic回归分析算法三组模型(ADC、T2WI、ADC+T2WI)在验证集的AUC分别为0.800、0.878、0.910,ADC+T2WI模型的AUC最大;应用R-Tree算法三种模型(ADC、T2WI、ADC+T2WI)在验证集的AUC分别为0.793、0.863、0.913,ADC+T2WI模型的AUC最大;两种机器学习算法构建的影像组学模型对肌层浸润性膀胱癌的诊断效能相当。结论基于Logistic回归分析法、R-Tree两种机器学习算法的影像组学模型对肌层浸润性膀胱癌均有较好的诊断价值,其中ADC+T2WI模型的诊断效能更优。Objective To explore the diagnostic value of machine learning-based MRI radiomics including Logistic regression and R-Tree algorithms in predicting muscle-invasive bladder cancer.Methods 132 patients with pathologically confirmed muscle-invasive(51)and non-muscle-invasive(81)bladder cancer underwent preoperative MRI including T2-weighted images(T2WI)and diffusion-weighted sequences.Volumes of interest were manually drawn on T2WI and apparent diffusion coefficient(ADC)maps by two radiologists using ITK-SNAP software.A list of radiomics features were extracted using the AK software,and the corresponding radiomics signature was constructed.Different machine learning was used to develop three prediction models of ADC,T2WI,and ADC+T2WI.The diagnostic efficiency of different models was evaluated by receiver operating characteristic(ROC)curve,and the areas under the curve(AUC),accuracy,specificity and sensitivity were obtained.Results Using the Logistic regression algorithm in the test set,the AUC of ADC+T2WI model(0.910)was higher than that of ADC(0.800)and T2WI(0.878).Using the R-Tree algorithm in the test set,the AUC of ADC+T2WI model(0.913)was also higher than that of ADC(0.793)and T2WI(0.863).The MRI radiomics model constructed by the two machine learning algorithms can predict muscle-invasive bladder cancer with similar diagnosis performance.Conclusion Both Logistic regression and R-Tree algorithms of machine learning-based MRI radiomics have good diagnostic value for predicting muscle-invasive bladder cancer,among which the ADC+T2WI model has the best diagnostic efficiency.

关 键 词：磁共振成像 肌层浸润性膀胱癌 影像组学 机器学习 

分 类 号：R445.2[医药卫生—影像医学与核医学] R737.14[医药卫生—诊断学]