基于机器学习的双参数MRI影像组学模型在移行带临床显著性前列腺癌预测中的价值  

The value of a machine learning-based biparametric MRI radiomics model in predicting clinically significant prostate cancer in the transitional zone

作　　者：李露 闫旭 马柯 王玉婷 金嵚 潘燚琪 孙琦[5] 麦筱莉 LI Lu;YAN Xu;MA Ke;WANG Yuting;JIN Qin;PAN Yiqi;SUN Qi;MAI Xiaoli(Department of Radiology,Nanjing Drum Tower Hospital Clinical College of Nanjing Medical University,Nanjing 210008,China;Department of Radiology,Xuancheng Central Hospital,Xuancheng,Anhui Province 242000,China;Department of Radiology,Nanjing Drum Tower Hospital Clinical College of Xuzhou Medical University,Nanjing 210008,China;Department of Radiology,Nanjing Drum Tower Hospital,the Affiliated Hospital of Medical School,Nanjing University,Nanjing 210008,China;Department of Pathology,Nanjing Drum Tower Hospital,the Affiliated Hospital of Medical School,Nanjing University,Nanjing 210008,China)

机构地区：[1]南京医科大学鼓楼临床医学院医学影像科,江苏南京210008 [2]宣城市中心医院影像科,安徽宣城242000 [3]徐州医科大学鼓楼临床学院医学影像科,江苏南京210008 [4]南京大学医学院附属鼓楼医院医学影像科,江苏南京210008 [5]南京大学医学院附属鼓楼医院病理科,江苏南京210008

出　　处：《实用放射学杂志》2024年第11期1837-1842,共6页Journal of Practical Radiology

摘　　要：目的基于机器学习的双参数磁共振成像(bpMRI)影像组学模型在移行带临床显著性前列腺癌(csPCa)预测中的价值。方法回顾性分析2家医疗中心共507例患者,所有患者在术前均接受前列腺MRI检查,病理学资料完整。病例分布为:csPCa 256例,非临床显著性前列腺癌(ciPCa)97例,前列腺良性增生(BPH)154例。使用R语言将中心1的数据按照7︰3随机分为训练组和测试组,将中心2的数据设为独立的外部验证组。分别提取T_(2)WI、扩散加权成像(DWI)的影像特征,并运用最小绝对收缩和选择算子(LASSO)降维筛选特征,分别构建基于T_(2)WI和T_(2)WI+DWI特征的2种数据集,使用随机森林(RF)、逻辑回归(LR)和支持向量机(SVM)建立6种预测模型。通过受试者工作特征(ROC)曲线及曲线下面积(AUC)、决策曲线分析(DCA)对比评估T_(2)WI特征、T_(2)WI+DWI联合特征的6种模型在前列腺疾病诊断中的效能。结果在训练组上进行特征筛选,T_(2)WI单序列、T_(2)WI+DWI双序列分别筛选出7、8个特征进行移行带csPCa预测,结果显示T_(2)WI+DWI双序列RF模型AUC效能最佳。训练组、测试组、验证组中的AUC分别为0.950、0.866、0.818,测试组准确性为0.805,敏感度为0.690,特异度为0.920;验证组准确性为0.726,敏感度为0.661,特异度为0.793。DCA显示在较宽的概率阈值范围内,T_(2)WI+DWI双序列RF模型的净获益最大。结论基于bpMRI影像组学模型可术前无创预测移行带csPCa,有助于临床制订诊疗决策。Objective To evaluate the value of a machine learning-based biparametric magnetic resonance imaging(bpMRI)radiomics model in predicting clinically significant prostate cancer(csPCa)in the transitional zone.Methods A retrospective analysis was conducted on 507 cases in two medical centers.All patients underwent prostate MRI examinations before surgery,with complete pathological data.The case distribution was as follows:256 cases of csPCa,97 cases of clinically insignificant prostate cancer(ciPCa),and 154 cases of benign prostatic hyperplasia(BPH).Using the R language,the data from Center One was randomly divided into training and test groups at a ratio of 7︰3,and the data from Center Two as an independent external validation group.The image features from T_(2)WI and diffusion weighted imaging(DWI)were extracted,and the least absolute shrinkage and selection operator(LASSO)was used to reduce dimensionality and filter features.Two datasets were constructed based on T_(2)WI features alone and combined T_(2)WI and DWI features.Six prediction models were established using random forest(RF),logistic regression(LR),and support vector machine(SVM).The efficacy of six models of T_(2)WI features and combined T_(2)WI and DWI features in the diagnosis of prostate diseases through receiver operating characteristic(ROC)curve,area under the curve(AUC),and decision curve analysis(DCA)were compared and evaluated.Results In the training group,feature screening identified 7 and 8 features from the T_(2)WI single sequence and the T_(2)WI with DWI dual sequence for csPCa prediction in the transitional zone.The results showed that the T_(2)WI with DWI dual sequence RF model had the highest AUC performance.The AUC of the training,test,and validation groups were 0.950,0.866,and 0.818,respectively.The test group accuracy was 0.805,sensitivity was 0.690,and specificity was 0.920;the validation group accuracy was 0.726,sensitivity was 0.661,and specificity was 0.793.DCA showed that within a wide probability threshold range,the T_(2)WI wit

关 键 词：前列腺癌 磁共振成像 纹理分析 影像组学 机器学习 

分 类 号：R737.25[医药卫生—肿瘤] R445.2[医药卫生—临床医学] TP181[自动化与计算机技术—控制理论与控制工程]