纯磨玻璃结节肺腺癌CT征象与其浸润程度的相关性分析及预测模型构建

Correlation analysis and prediction model construction of CT signs and infiltration degree of pure ground-glass nodule lung adenocarcinoma

摘要:

目的 探究纯磨玻璃结节（pure ground-glass nodule，pGGN）肺腺癌计算机断层扫描（computed tomography，CT）征象与其浸润程度的相关性，建立CT征象与浸润程度的预测模型。方法 回顾性分析424例经手术切除、病理活检证实且胸部CT表现为pGGN的肺腺癌患者临床资料及CT征象，根据病理活检结果分为非典型腺瘤样增生、原位腺癌、微浸润腺癌和浸润性腺癌4组，对组间差异采用卡方检验或Fisher确切概率法进行统计分析。对有统计学意义的结果，使用怀卡托智能分析环境（Waikato environment for knowledge analysis，WeKa）中的6种学习算法进行预测模型构建，并验证准确性，挑选出最适用于本研究的预测模型。结果 4组间在结节直径、结节密度值上的差异具有统计学意义（P<0.001），对应的直径平均值分别为6.90、8.65、10.71、14.56 mm，对应的密度平均值分别为-633.16、-543.04、-401.03、-322.94 HU，随着病灶的浸润程度加重，结节的直径及密度值呈现明显的上升趋势。4组间在结节边界、分叶、毛刺、血管穿行、胸膜凹陷、空气支气管征、空泡征等的差异具有统计学意义（P<0.05），而结节的生长位置，年龄、吸烟史、直系亲属肺癌家族史等差异无统计学意义（P>0.05）。随机森林算法所构建的模型预测准确率为76.42%~79.72%，Kappa系数为0.597~0.670，受试者工作特征（receiver operating characteristic，ROC）曲线下面积均大于0.9，在误差指标中表现最优，是最适合于本研究的预测模型。结论 pGGN的不同CT征象与其浸润程度密切相关，可以用于建立预测模型。基于随机森林算法所建模型，在有创干预前早期快速识别pGGN浸润程度的平均准确率为78.07%，准确度最高，对肺癌预测具有潜在应用价值。

Abstract:

Objective To explore the correlation between computed tomography（CT） signs and infiltration degree of pure ground-glass nodule（pGGN） lung adenocarcinoma，and establish a prediction model of CT signs and infiltration degree.Methods The clinical data and CT signs of 424 patients with lung adenocarcinoma confirmed by surgical resection，pathological biopsy and chest CT findings of pGGN were analyzed retrospectively，and according to the results of pathological biopsy，they were divided into four groups： atypical adenomatous hyperplasia，adenocarcinoma in situ，minimally invasive adenocarcinoma and invasive adenocarcinoma. The Chi-square test or Fisher’s exact probability test was used for statistical analysis of differences between groups. For the results with statistical significance，six learning algorithms in Waikato environment for knowledge analysis（WeKa） were used to build the prediction model，verify the accuracy，and select the prediction model most suitable for this study.Results The differences in nodule diameter and nodule density between the four groups were statistically significant（P<0.001），with the corresponding mean diameters of 6.90，8.65，10.71 and 14.56 mm，and the corresponding mean densities of -633.16，-543.04，-401.03 and -322.94 HU，respectively. The diameter and density of the nodules showed an obvious upward trend with the increase of the degree of invasion of the lesions. There were statistically significant differences among the four groups in nodule boundary，lobulation，burr，vascular perforation，pleural indentation，air bronchogram sign，and vacuole sign（P< 0.05），but there were no statistically significant differences in nodule growth position，age，smoking history，family history of lung cancer in immediate family（P>0.05）. The prediction accuracy of the model constructed by the random forest algorithm fluctuated between 76.42% and 79.72%，the Kappa coefficient fluctuated between 0.597 and 0.670， and the area under the receiver operating characteristic（ROC） curve was greater than 0.9，which was the best among the error indicators. Thus，it is the most suitable prediction model for this study.Conclusion The different CT signs of pGGN are closely related to the degree of infiltration， and can be used to establish a prediction model. Based on the model built by random forest algorithm，the average accuracy of quickly identifying the pGGN infiltration degree in the early stage before invasive intervention is 78.07%，which is the highest accuracy，and can potentially be used in the prediction of lung cancer in the future.