影像诊断淋巴结包膜外侵犯的影像组学模型构建及对前列腺癌侵袭性的预测

Constructing radiomics model of radiologic extranodal extension and predicting aggressiveness of prostate cancer

目的构建影像诊断淋巴结包膜外侵犯(rENE)的影像组学模型,探讨其预测前列腺癌(PCa)侵袭性的价值。方法回顾性收集经病理证实为PCa,且术后病理或MRI提示肿瘤为N1期的病人160例,平均年龄(68.6±9.19)岁。所有病人术前均进行了双参数(T_(2)WI+ADC)MRI检查。根据原发病灶的侵袭性,将病人按照国际泌尿病理协会(ISUP)分级分组分为ISUP≥4组(124例)和ISUP<4组(36例)。使用随机分层抽样法将纳入病例以7∶3的比例分为训练集(112例)及测试集(48例)。采用5折交叉验证构建交叉验证集。由放射医师判定病人的影像rENE状态。采用FAE软件提取转移淋巴结的影像特征,选取8种机器学习算法构建模型,采用受试者操作特征(ROC)曲线下面积(AUC)评估各影像组学模型的诊断效能,以AUC值随特征数变化的波动程度评价稳定性。综合分析AUC值及模型稳定性筛选最优影像组学模型。ISUP≥4组和ISUP<4组间、训练集及测试集间一般资料的比较采用t检验、Mann-Whitney U检验和卡方检验。采用多因素Logistic回归分析ISUP≥4组的独立预测因素。结果ISUP≥4组的总前列腺特异性抗原(tPSA)、rENE+占比均高于ISUP<4组(均P<0.05);ISUP≥4组中T4期的占比最高(49.2%),ISUP<4组中T_(2)期的占比最高(30.5%)(均P<0.05)。训练集与测试集间,一般资料差异均无统计学意义(均P>0.05)。最终纳入由2个特征构建的Zscore_PCC_ANOVA_2_NB模型。该模型诊断rENE的AUC值在训练集、测试集及交叉验证集中均较高,分别为0.952、0.960和0.954。在训练集、测试集及全部病例中,影像组学模型预测ISUP≥4组的AUC值分别为0.862、0.944和0.923。多因素Logistic回归分析显示Zscore_PCC_ANOVA_2_NB模型及较高的tPSA值为ISUP≥4的独立预测因素。结论rENE的影像组学模型对PCa的病理分级预测效能较高,可无创、定量预判PCa侵袭性。

Objective To construct radiomics models of radiologic extranodal extension(rENE),and evaluate their value in predicting the aggressiveness of prostate cancer(PCa).Methods Patients with pathological N1 staging or MRIdiagnosed N1 staging were collected retrospectively,a total of 160 patients were collected with a mean age of 68.6±9.19 years old.All the patients underwent bi-parameter(T_(2)WI+ADC)MRI examination.According to the aggressiveness of the primary lesion,patients were divided into International Society of Urological Pathology(ISUP)grades≥4 group(124 patients),and ISUP<4 group(36 patients).Randomized stratified sampling method was used to further divide the patients into a training set(112 patients)and a test set(48 patients)in a ratio of 7∶3.The cross-validation set was constructed by using 5-fold crossvalidation.The rENE status was evaluated by radiologists.The FeAture Explorer Pro was used to extract the radiomics features of the metastatic lymph nodes,and eight machine learning algorithms were used to construct the radiomic models.The diagnose ability of each model was tested by receiver operating characteristic(ROC)curve and the areas under the curve(AUC).The stability was evaluated by the degree of fluctuation of AUC value with the change of feature number.Stability and AUC values were combined to select the best model.The t-test,Mann-Whitney U test,and chi square test were used to compare general data between the ISUP≥4 and ISUP<4 groups,and between the training and test sets.Multivariate Logistic regression was used to determine the independent predictors of the ISUP≥4 group.Results The tPSA and rENE+percentage were significantly higher in the ISUP<4 group(all P<0.05),the majority of T stage in the ISUP<4 group was T_(2)(30.5%),and T4(49.2%)in the ISUP≥4 group(both P<0.05).No significant differences existed between the data of the train and test groups(all P>0.05).The best model selected was Z score_PCC_ANOVA_2_NB,combined by 2 radiomics features.AUC values were higher in the training set,test set,and cross-validation set for diagnosing rENE,which were 0.952,0.960,and 0.954 respectively.The AUC of the model for predicting ISUP≥4 grading patients in the training set,test set,and all patients were 0.862,0.944,and 0.923 respectively.Multivariate Logistic regression indicated that,the Zscore_PCC_ANOVA_2_NB model and high tPSA value were the independent factors of ISUP≥4 group patients.Conclusion The rENE radiomics model could predict highly aggressive PCa non-invasively and quantitatively.