基于影像组学的肺亚实性结节侵袭性预测模型建立及分析

Establishment and analysis of prediction model for invasive subsolid pulmonary nodules based on radiomics

目的基于影像组学特征构建预测模型,预测肺亚实性结节非侵袭性/侵袭性的病理亚型。方法回顾性收集2015年1月至2019年9月东南大学附属中大医院及东部战区总医院胸部高分辨率计算机断层扫描(HRCT)表现为肺亚实性结节、手术病理结果为不典型瘤样增生(AAH)、原位腺癌(AIS)、微浸润性腺癌(MIA)、浸润性腺癌(IA)共352例患者资料,其中男108例,女244例,年龄[M(Q1,Q3)]57(50,65)岁。根据病理分为非侵袭组233例和侵袭组119例。按照训练集:内部测试集:外部测试集大约3∶1∶1的比例分为训练集(215例,非IA/IA为155例/60例)、内部测试集(69例,非IA/IA为52例/17例)及外部测试集(68例,非IA/IA为26例/42例,均为东部战区总医院病例)。记录特定的结节定量参数、组学特征、形态学特征、患者临床资料、血清肿瘤标志物。LASSO回归用于构建组学标签。使用logistic回归分析分别构建形态学模型、CT模型、综合模型,在测试集进行验证。结果基于训练集筛选出2个最有意义的特征为Shape_MinorAxis(Gradient)、Glszm_ZoneEntropy(LBP)(均P<0.001),构建组学标签=1.06575×Shape_MinorAxis(Gradient)+0.03058×Glszm_ZoneEntropy(LBP)。综合组学标签、胸膜凹陷征、定量参数(直径、平均密度)构建的CT模型为最优模型,回归方程Ln(P/1⁃P)=-2.41711+1.03160×组学标签+1.20306×直径+1.61421×(胸膜凹陷征=有)在训练集、测试集的AUC分别为0.954(95%CI:0.927~0.981)、0.865(95%CI:0.764~0.966),优于形态学模型0.857(95%CI:0.796~0.918)、0.818(95%CI:0.686~0.949)及综合模型0.951(95%CI:0.921~0.981)、0.856(95%CI:0.730~0.982)。结论综合构建的CT模型对预测以亚实性结节为表现的侵袭性肺腺癌具有较好的预测效能。

Objective To evaluate the best radiomic features based prediction model for identifying the histopathological subtypes of invasive adenocarcinoma or noninvasive pulmonary nodules appearing as subsolid nodules.Methods A total of 352 patients(108 males and 244 females,median age was[M(Q1,Q3)]57(50,65),underwent high‑resolution chest CT and appearing as subsolid nodules and further treated by surgical resection whose subsequently pathological results were classified as atypical adenomatous hyperplasia(AAH),carcinoma in situ(AIS),microinvasive carcinoma(MIA),invasive adenocarcinoma(IA),from January 2015 to September 2019,in Radiology Department of Zhongda Hospital Affiliated to Southeast University and Jinling Hospital,Medical School of Nanjing University were retrospectively collected.They were divided into non‑invasive group(n=233)and invasive group(n=119)according to pathological findings.According to the ratio of training set:internal test set:external test set,which is about 3∶1∶1,the patients in Zhongda Hospital Affiliated to Southeast University were randomly divided into training set(n=215,non‑IA∶IA 155∶60)and internal test set(n=69,non‑IA∶IA 52∶17),meanwhile a certain number of patients in Jinling Hospital,Medical School of Nanjing University(n=68,non‑IA∶IA 26∶42)were randomly selected as an independent external test set.Particular quantitative parameters of the nodules,radiomic features,morphological characteristics,clinical data,and serum tumor markers were recorded.Radiomic label was constructed using LASSO regression method.The morphological model,CT model and comprehensive model were constructed by binary logistic regression and were verified in test sets,respectively.Results Shape_MinorAxis(Gradient),Glszm_ZoneEntropy(LBP)were selected as the two most significant features based on training set.Radiomic tag=1.06575×Shape_MinorAxis(Gradient)+0.03058×Glszm_ZoneEntropy(LBP).Comparing the prediction performance of all models in each data cohort,the CT model(Ln(P/1‑P)=-2.41711+1.03160×Radimic tag+1.20306×Diameter+1.61421×(Pleural indentation sign=Y)constructed by radiomic label,pleural depression,and quantitative parameters(diameter,average density)was much better than other models and was chosen as the optimal model,with an AUC of CT models in training cohort and test cohort was 0.954(95%CI:0.927⁃0.981),0.865(95%CI:0.764⁃0.966),better than morphological model 0.857(95%CI:0.796⁃0.918),0.818(95%CI:0.686⁃0.949)and comprehensive model 0.951(95%CI:0.921⁃0.981),0.856(95%CI:0.730⁃0.982),respectively.Conclusion The integrative CT model has a better prediction efficiency for identifying invasive or noninvasive nodules appearing as subsolid nodules.