基于肺癌CT的决策树模型在肺癌诊断中的应用
Application of decision tree model combined with CT in diagnosis of lung cancer

摘要目的：探讨基于肺癌CT的决策树模型在肺癌辅助诊断中的意义。方法：观察61例肺癌患者和50例肺部良性疾病患者的CT片，对边缘情况、毛刺征、小泡征等放射学特征进行量化，结合临床资料对肺癌相关的危险因素行单因素非条件logistic回归和多因素非条件logistic逐步回归分析，采用有意义的因素分别建立决策树、logistic回归2种肺癌诊断模型。采用筛检实验评价指标和ROC曲线比较2种模型对预测集样本的预测效果。结果：多因素非条件logistic逐步回归分析显示年龄、边缘情况、毛刺征、小泡征及纵隔肿大淋巴结对鉴别诊断肺癌意义较大；logistic回归模型对预测集的预测灵敏度、特异度、准确度、阳性预测值和阴性预测值分别为86.89%、84.00%、85.59%、86.89%、84.00%，决策树模型上述指标分别为95.08%、86.89%、90.99%、89.23%、93.48%；logistic回归模型和决策树模型ROC曲线下面积分别为0.854(0.775~0.914)和1.000(0.967~1.000)，差异有统计学意义（Z=4.273，P<0.05）。结论：建立的决策树模型诊断肺癌的效果优于logistic回归模型。
AbstractAim: To distinguish lung cancer patients and patients with benign lung diseases by using decision tree model combined with CT. Methods: Radiologists extracted edge features, speculations, small bubble sign and other CT radiology features after they carefully observed the CT results of 61 patients with lung cancer and 50 patients with benign lung diseases. Then radiologists quantified the score of the 15 extracted radiology features. Unvaried and multivariate logistic regression analysis with clinical parameters was applied to filter the meaningful factors with lung cancer risk and then logistic regression model and decision tree model were developed. Factors of screening and ROC curves were used for evaluating the effect of two models in predicting prediction set.Results: The multivariate logistic regression analysis showed that age, edge features, speculations, small bubble sign and mediastinal lymph nodes had greater significance for differential diagnosis of lung cancer. The sensitivity, specificity, accuracy, positive prognostic value, and negative prognostic value for logistic regression model were 86.89%, 84.00%, 85.59%, 86.89% and 84.00%, respectively, and those for decision tree model were 95.08%, 86.89%, 90.99%, 89.23%, 93.48%, respectively. Areas under the ROC curve were 0.854(0.775-0.914) and 1.000(0.967-1.000) for logistic regression model and decision tree model, respectively（Z=4.273，P<0.05）. Conclusion: Decision tree model is better than logistic regression model in diagnosis of lung cancer