分类决策树辅助CT诊断孤立性肺结节的方法学研究

Classification decision tree in CT imaging： application to the differential diagnosis of solitary pulmonary nodules

目的应用分类与回归决策树（CART）算法构建CT显像鉴别良恶性孤立性肺结节（SPN）预测模型，探讨数据挖掘技术在SPN影像诊断中的应用价值。方法分别提取12个临床指标和22个CT征象指标作为CART预测SPN良恶性的输入指标。连续性纳入自2003年7月至2006年7月间经病理证实的SPN，且术前行CT检查的患者116例，其中良性结节62例，恶性结节54例。采用CART建立用于预测良恶性SPN的分类决策树模型，并通过交互印证的方法计算该模型的诊断准确性。同时设低年资医师诊断组和高年资医师诊断组，采用盲法进行独立阅片判断SPN的良恶性。采用受试者操作特征（ROC）曲线比较3组间的诊断效能。结果（1）成功建立了能够判断SPN良恶性的CART诊断模型，其中含有8条诊断规则，最低相对错误代价为0．199，CART对SPN具有决策意义的最重要的前3位决策指标为结节的毛刺征、患者年龄和病灶部位。（2）CART、高年资医师和低年资医师对SPN良恶性诊断的ROC曲线下面积分别为0．910±0．029、0．827±0．038、0．612±0．052。CART与低年资医师ROC曲线下面积差（DBF）=0．297，P〈0．01；与高年资医师DBF=0．083，P〈0．05；高年资医师与低年资医师DBF=0．214，P〈0．01。CART诊断效能高于高年资医师和低年资医师，高年资医师高于低年资医师。结论CART是具有强大学习能力的数据挖掘工具，可以对SPN的良恶性进行正确判断，为实现人工智能在影像诊断中的应用提供重要的方法学依据。

Objective To establish classification and regression tree （CART） for differentiating benign from malignant solitary pulmonary nudules （SPN）. Methods One hundred and sixteen consecutive cases with 116 solitary pulmonary nodules, which finally were pathologically proven 54 malignant nodules and 62 benign nodules, were prospectively registered in this research. Twelve clinical presentations and 22 CT findings were collected as predictors. A classification tree was established to distinguish benign SPNs from malignant ones. In the observer test, two groups （one made of junior radiologists and one of senior radiologists） were independently presented with clinical information and CT images without knowing the pathologic and machine-learning results. Performance of observers and CART were compared by receiver operating characteristic analysis. Results Receiver operating characteristic analysis showed areas under the curve of CART, senior radiologists and junior radiologists respectively were 0. 910 ± 0. 029, 0. 827 ± 0. 038, 0. 612 ±0. 052. Difference between areas（DBF） between CART and junior radiologists was 0. 297（P 〈0. 01 ）. DBF between CART and senior radiologists was 0. 083 （ P 〈0. 05 ）. DBF between senior and junior radiologists was 0. 214 （P 〈0. 01）. CART showed a best diagnostic efficiency, followed by junior radiologists, and then senior radiologists. Conclusion Our data mining techniques using CART prove a high accuracy in differentiating benign from malignant pulmonary nodules based on clinical variables and CT findings. It will be a potentially useful tool in further application of artificial intelligence in the imaging diagnosis.