高分辨CT影像组学模型鉴别亚厘米肺纯磨玻璃结节腺体前驱病变与微浸润腺癌的价值

Value of high-resolution CT radiomics model in differentiating glandular precursor lesions and minimally invasive adenocarcinoma presenting as subcentimeter pure ground glass nodules

目的 探讨基于高分辨CT影像组学模型鉴别表现为亚厘米肺纯磨玻璃结节的腺体前驱病变和微浸润腺癌(MIA)的价值。方法 回顾性分析2020年7月~2022年4月经手术病理证实的亚厘米纯磨玻璃结节患者共计68例(75个肺结节),包括6个非典型腺瘤样增生、26个原位癌及43个MIA,根据病理类型分为腺体前驱病变组(非典型腺瘤样增生+原位癌)和微浸润组(MIA),将其分为训练组54例(60个pGGN),验证组14例(15个pGGN)。采集临床资料(年龄、性别)、CT定性参数(边界、毛刺、分叶、支气管异常征、内部血管征、空泡征、胸膜牵拉征)及定量参数(最长径、最短径、平均CT值、最大CT值、最小CT值)。利用ITK-SNAP软件对每个纯磨玻璃结节行手动分割并导入AK软件进行影像特征提取。采取单因素及多因素分析方法筛选出训练组中两亚组之间差异有统计学意义的变量,利用多元Logistic回归的方法构建影像组学模型、临床特征模型及联合模型。通过ROC曲线及计算曲线下面积(AUC)对各模型的预测效能进行比较,使用Delong’s检验比较各模型之间的差异。采用校正曲线及决策曲线分析评估联合模型的校准度及临床应用性,采用Hosmer-Lemeshow检验分析联合模型预测值与观测值之间的拟合程度。结果 联合模型在训练组和验证组中均具有最高的诊断效能(训练组AUC=0.857,95%CI:0.764~0.951,P<0.0001;验证组AUC=0.84,95%CI:0.592~1.000,P=0.0071),高于影像组学模型(训练组AUC=0.835,95%CI:0.735~0.935,P<0.0001;验证组AUC=0.82,95%CI:0.563~1.000,P=0.0145)和临床特征模型(训练组AUC=0.764,95%CI:0.636~0.864,P<0.0001;验证组AUC=0.63,95%CI:0.347~0.913,P=0.3677)。联合模型在训练组和验证组中的预测观察值与实际观察值之间均具有良好的一致性。结论 基于高分辨CT影像组学和临床特征构建的联合模型有助于术前鉴别表现为亚厘米肺纯磨玻璃结节的腺体前驱病变和MIA,提升肺结节诊治及管理水平。

Objective To assess the efficacy of a radiomics model rooted in high-resolution CT imaging for the differentiation of precursor glandular lesions and minimally invasive adenocarcinoma(MIA) manifesting as subcentimeter pure ground-glass nodules(pGGN).Methods A total of 68 patients(75 pulmonary nodules) with subcentimeter pGGN confirmed by surgical pathology from July 2020 to April 2022 were retrospectively analyzed,including 6 atypical adenomatous hyperplasia(AAH),26 adenocarcinoma in situ(AIS) and 43 MIA.According to the pathological type,the patients were divided into precursor glandular lesions group(AAH+AIS) and minimally invasive group(MIA),including 54 cases in the training group(60 pGGN)and 14 cases in the validation group(15 p GGN).Clinical data(age,gender),CT qualitative parameters(margin,spiculation,lobulation,air bronchogram,internal vseesl sign,bubblen,pleural attachment) and quantitative parameters(longest diameter,shortest diameter,average CT value,maximum CT value,minimum CT value) were collected.Manual segmentation of each pGGN was performed using ITK-SNAP software,and image features were extracted using AK software.Statistical analyses included univariate and multivariate methods to identify significant differences between the two subgroups in the training group.We used these analyses to create imaging radiomics models,clinical models,and combined models through multivariate Logistic regression.The prediction efficiency of each model was compared by ROC curve and the area under the curve(AUC),and Delong's test was used to compare whether there were significant differences among the models.The calibration curve and the decision curve analysis were used to evaluate the calibration and clinical application of the combined model,and Hosmer-Lemeshow test was used to analyze the fitting degree between the predicted value and the observed value of the combined model.Results The combined model had highest diagnostic efficiency in both the training group and the text group(AUC=0.857,95% CI:0.764-0.951,P<0.0001 in the training group;AUC=0.84,95% CI:0.592-1.000,P=0.0071 in the text group),which was higher than the radiomics model(AUC=0.835,95% CI:0.735-0.935,P<0.0001 in the training group;AUC=0.82,95% CI:0.563-1.000,P=0.0145 in the text group) and clinical model(AUC=0.764,95% CI:0.636-0.864,P<0.0001 in the training group;AUC=0.63,95%CI:0.347~0.913,P=0.3677 in the text group).Furthermore,the combined model demonstrated a commendable degree of consistency between its predicted values and actual observations in both the training and text group.Conclusion The combined model based on CT radiomics and clinical features is helpful to distinguish precursor glandular lesions and MIA which presenting as subcentimeter pure ground glass nodules before operation,and improve the level of diagnosis,treatment and management of pulmonary nodules.