基于机器学习探究临床联合增强CT影像组学特征对肺结节良恶性的鉴别诊断价值

A Machine Learning-Based Approach Identifying the Value of Clinical and Enhanced CT Imaging Histological Features for the Diagnosis of Benign and Malignant Pulmonary Nodules

目的:通过机器学习方式对肺结节的增强CT特征提取影像组学和临床特征分析对其定性诊断的价值。方法:回顾性分析本院2018年6月至2022年7月有明确手术标本病理证实的89例肺结节患者,其中良性结节患者37例,恶性结节患者52例,按照8∶2比例将其分为训练集和测试集,逐层提取患者平扫期、动脉期、静脉期病灶区域的感兴趣区(ROI),采用软件提取影像组学纹理特征,使用单因素分析、多因素分析和最小绝对收缩和选择算子进行特征筛选,并采用多种机器学习的方法建立预测其良恶性的模型,基于特征权重建立SHAP值图,最后使用DCA曲线分析患者获益程度。结果:训练集72例,测试集17例,共提取了影像组学特征和临床特征数量分别为2800和31个。经特征筛选,共保留13个影像特征和4个临床特征。临床特征中既往肺部基础疾病史、细胞角蛋白19的可溶性片段(CYFRA21-1)及影像特征中的毛刺征、分叶征在良恶性组中差异有统计学意义(P<0.05),采用了多种机器学习算法,其中效能最高的是XGBoost算法,DCA结果表明患者可获得良好收益。结论:基于增强CT及肿瘤标志物等建立的XGBoost模型对鉴别肺结节的性质具有重要价值。

Objective To identify the value of clinical histological features and the enhanced CT imaging features in the di⁃agnosis of pulmonary nodules by machine learning.Methods A retrospective study was conducted for 89 patients with pul⁃monary nodules confirmed by surgical specimens in the Xiangyang No.1 People′s Hospital from June 2018 to July 2022,in⁃cluding 37 patients with benign nodules and 52 patients with malignant nodules.The patients were classified into a training set and a validation set at a ratio of 8∶2.The regions of interest(ROI)in the lesion was extracted during the plain scan,arterial phase and venous phase.The imaging features were extracted by software and screened by univariate analysis,mult⁃ivariate analysis and least absolute shrinkage and selection operator.Furthermore,machine learning methods were employed to establish the model for predicting the benign or malignant nodules,and the diagram of weighted SHapley Additive exPla⁃nation(SHAP)values was established.Finally,the decision curve analysis(DCA)was employed to analyze the patient benefits.Results The training set and validation set included 72 and 17 patients,respectively.A total of 2800 imaging fea⁃tures and 31 clinical features were extracted.After screening,13 imaging features and 4 clinical features were retained.The clinical features of history of underlying lung diseases and cytokeratin 19 fragment antigen 21-1(CYFRA 21-1),as well as the imaging features of burr sign and lobar sign, showed significant differences between the benign and malignant groups (P<0.05). Among the various machine learning methods, XGBoost demonstrated the highest performance. The DCA results indicated good patient benefits. Conclusion The XGBoost model, based on enhanced CT and tumor markers, is of great value in identifying the nature of pulmonary nodules.