论文部分内容阅读
目的利用神经网络分析法构建肺磨玻璃密度结节(GGN)侵袭性的CT预测模型,探讨其预测的准确性。方法回顾性分析203例经手术病理证实为肺腺癌的肺GGN的CT影像特征。采集患者基本信息,统计肺结节密度(纯磨玻璃结节或混合磨玻璃结节)、是否含有内核、大小、实性成分比例;采用评分法对空泡征、胸膜牵拉征、血管集束征三个影像特征进行量化评分,利用单因素方差分析各CT特征在不同病理分组间的差异,利用神经网络法将病例随机分为培训组(103例)和检验组(100例),建立各CT特征与GGN病理之间的预测模型。结果203例肺GGN中AAH 20例,AIS 26例,MIA 74例,I-ADC 83例。四组病理类型间的结节性质、直径、实性成分比例以及三个影像特征通过单因素方差分析均存在显著性差异(P<0.05)。基于此数据而使用神经网络的“多层感知器”(MLP)建立预测模型。培训组总体预测准确率为80.6%(AAH 92.9%,AIS 38.5%,MIA 91.2%,I-ADC81.0%)检验组预测总体准确率为72.0%(AAH 50.0%,AIS 46.2%,MIA 72.5%,I-ADC 82.9%),各自变量在模型中的重要性WTMW/WTLW(0.270,100%),影像特征评分(0.263,97.6%),WTMW(0.099,36.7%),WTLW(0.097,36.0%),胸膜牵拉征(0.085,31.5%),血管集束征(0.084,31.0%),空泡征(0.051,18.8%),内核(0.027,9.9%),结节密度(0.025,9.4%)。结论基于神经网络建立的GGN侵袭性CT预测模型可用于GGN病理侵袭性评估。
Objective To construct an invasive CT predictive model of lung-worn glass density nodules (GGN) by using neural network analysis and investigate its accuracy. Methods The CT features of lung GGN in 203 patients with lung adenocarcinoma confirmed by pathology were retrospectively analyzed. The basic information of the patients was collected, and the density of pulmonary nodules (purely ground glass nodules or mixed ground glass nodules) was counted, and the proportion of kernel, size and solid component was counted. The scoring method was used to evaluate the characteristics of vacuole, pleural traction, The three imaging features were quantified. One-way analysis of variance (ANOVA) was used to analyze the differences of CT features among different pathological groups. The cases were divided into training group (n = 103) and testing group (n = 100) Prediction model between CT features and GGN pathology. Results In the 203 cases of lung GGN, 20 cases were AAH, 26 cases were AIS, 74 cases were MIA and 83 cases were I-ADC. There was a significant difference (P <0.05) in nodule quality, diameter, proportion of solid components and three image features among four groups by one-way ANOVA. Based on this data, a predictive model was built using the Neural Layer’s “Multilayer Perceptron” (MLP). The overall prediction accuracy of the training group was 72.0% (AAH 50.0%, AIS 46.2%, MIA 72.5%) with the overall prediction accuracy of 80.6% (AAH 92.9%, AIS 38.5%, MIA 91.2%, and I-ADC 81.0% , I-ADC 82.9%), the importance of each variable in the model WTMW / WTLW (0.270, 100%), imaging feature score , Pleural traction sign (0.085, 31.5%), vascular bundle sign (0.084, 31.0%), vacuole sign (0.051, 18.8% . Conclusion The GGN invasive CT prediction model based on neural network can be used to assess the pathological invasion of GGN.