影像组学结合机器学习在鉴别脊柱结核与转移瘤中的价值研究

Machine Learning in Differentiating Spinal Tuberculosis and Metastasis Spine Tumors-Based on MR Radiomics

目的分析脊柱结核(TBS)与脊柱转移瘤(MST)患者受累椎体的MRI表现与影像组学特征,评价基于MRI特征的逻辑回归(Logistic)模型与基于影像组学特征的机器学习模型的诊断效能。方法搜集本院经病原学检查与手术病理证实的TBS与MST患者,利用Logistic分析MRI特征。提取受累椎体T_(2)WI脂肪抑制序列(T_(2)-FS)的影像组学特征,组内相关系数(ICC)评价组学特征值测量的可重复性。依次使用t检验、SelectKBest以及最小绝对收缩和选择算子(LASSO)筛选特征。利用交叉验证(CV)划分数据集,随机森林(RF)及支持向量机(SVM)模型在训练集上进行监督学习,在测试集上进行评价,并与Logistic模型相比较。受试者工作特征(ROC)曲线及决策曲线分析(DCA)分别用来评价机器学习的分类效能及实际临床净收益,校准曲线评估模型的预测误差,Z检验用来比较ROC曲线下面积(AUC)之间的差异。结果101例患者被纳入样本(51例TBS,50例MST),筛选到3个MRI特征构建Logistic模型,6个影像组学特征(ICC均>0.75)进行机器学习。训练集上RF的AUC为0.997(95%CI 0.994~1.000),SVM为0.991(95%CI 0.981~1.000),差异无统计学意义(P=0.250)。Logistic的AUC为0.871(95%CI 0.800~0.941),低于测试集上的RF:0.993(95%CI 0.986~1.000)与SVM:0.989(95%CI 0.979~1.000)(P均<0.05)。DCA表明,RF的净获益优于SVM,优于Logistic。校准曲线显示三个模型预测概率与真实概率间的差异无统计学意义(P均>0.05),预测误差均值分别为0.019(RF)、0.019(SVM)、0.052(Logistic)。结论基于受累椎体的影像组学特征进行机器学习鉴别TBS与MST是可行的,其结果优于基于MRI特征的Logistic模型,对于术前减少侵入性检查与指导治疗有着重大意义。

Objective To analyze MRI performance and radiomics features constructed from involved centrums of patients with tuberculous spondylitis(TBS)and metastatic spine tumors(MST),and evaluate the efficiency of logistic model based on MRI features and machine learning models based on radiomics in discriminating among them.Methods Retrospectively collected patients who confirmed by surgical pathological biopsy.T_(2) fat-suppression images before surgery were used to extract features,intraclass correlation coefficient(ICC)was used to assess the repeatability of results.T-test,Select-Kbest and least absolute shrinkage and selection operator(LASSO)were used successively to select features.Cross-validation(CV)was used to divide the data set.Support vector machines(SVM)and random forest(RF)were trained on the training set and evaluated on the test set.The models’efficiency of classification and clinic benefit were evaluated by receiver operating characteristic(ROC)curves and decision curve analysis(DCA).The calibration curve evaluated the prediction error of the model.The statistical differences of AUC used Z test to verify.Results 101 patients were included finally,51 TBS and 50 MST.3 MRI features and 6 radiomics features with ICC over 0.75 were retained.In training set,area under curve(AUC)of ROC had no statistically significant differences in RF and SVM,P=0.250.The AUC of Logistic were 0.871(95%CI:0.800-0.941),lower than RF 0.993(95%CI:0.986-1.000)and 0.989(95%CI:0.979-1.000)in test set,P values were all less than 0.05.DCA showed that the order of clinical net benefit was RF,SVM and logistic.The calibration curve showed that the difference of the three models between the predicted probabilities and the true probabilities was not statistically significant,P value were all greater than 0.05,and the average prediction errors respectively were 0.019(RF),0.019(SVM),and 0.052(logistic).Conclusions SVM and RF models showed excellent capacities of classification in TBS and MST,the results were better than the logistic model based on MRI features.Machine learning has significance for reducing invasive examinations and guiding treatment before surgery.