Abstract:

Objective To discuss the imaging risk factors for invasiveness of mixed ground-glass nodule （mGGN）-like pulmonary adenocarcinoma， to construct a machine learning model combining clinical imaging features and radiomics， and to clarify its diagnostic efficacy. Methods A total of 261 patients with postoperative pathologically confirmed mGGN-like pulmonary adenocarcinoma were retrospectively collected， and 278 mGGNs met the inclusion criteria. According to the pathological results， they were divided into indolent group （n=73） and invasive group （n=205）. The clinical data and imaging data of the patients in two groups were collected. All the patients underwent thin-slicecomputed tomography（CT） scanning； radiomic features were extracted. Feature dimensionality reduction was performed on the data via Z-score standardization combined with LASSO regression， and a 10-fold cross-validation was adopted to complete the screening process. The optimal omics features were identified， and a radiomics model based on the machine learning algorithm of Logistic Regression （LR） was constructed. Univariate and multivariate logistic regression analyses were used to identify the imaging features with significance for invasiveness of mGGN-like pulmonary adenocarcinoma； nomogram was drawn by combining the optimal radiomic features； the area under the receiver operating characteristic （ROC） curve （AUC）， calibration curve， and decision curve analysis （DCA） were used to evaluate the predictive performance of the model. Bootstrap method was used for model validation. Results Among the 278 included mGGNs， a total of 205 patients were pathologically confirmed as invasive mGGN-like pulmonary adenocarcinoma. Compared with indolent group， the optimal radiomic features extracted from the patients in invasive group included maximum three dimension（3D） diameter， the 90th percentile， inverse difference moment normalized， and surface area to volume ratio. The univariate Logistic analysis results showed that the incidences of spiculation， lobulation， air bronchogram sign， pleural indentation sign， spinous process， long diameter， short diameter， and mean CT value in the patients in invasive group were higher than those in indolent group （P<0.05）. The multivariate Logistic regression analysis results showed that lobulation was significantly associated with the risk of group classification ［odds ratio（OR）=3.724， 95% confidence interval（CI）： 1.373-10.100， P=0.010］； for every 1 mm increase in long diameter， the risk of group classification was increased by 31.8% （OR=1.318， 95%CI： 1.152-1.509， P<0.001）； for every 1 HU increase in mean CT value， the risk of invasive carcinoma was increased by 0.6% （OR=1.006， 95%CI： 1.003-1.009， P<0.001）. The AUC of the fusion nomogram constructed by combining lobulation， nodule long diameter， mean CT value and radiomics score was 0.858 （95%CI： 0.813-0.914）， and the bootstrap concordance index （C-index） of model was 0.835. The DCA curve results showed that the model had good clinical applicability. The Hosmer-Lemeshow test results showed that the nomogram had a good fit. Conclusion Lobulation， nodule long diameter and mean CT value are the risk factors for invasiveness of mGGN-like pulmonary adenocarcinoma. The fusion nomogram has good diagnostic efficacy for invasiveness of mGGN-like pulmonary adenocarcinoma， and may provide a new method for non-invasive diagnosis and individualized treatment in clinical practice.

Key words: Mixed ground-glass nodules, Lung neoplasm, Early screening for lung cancer, Radiomics, Risk factor, Risk prediction model