Abstract:

Objective To differentiate minimally invasive adenocarcinoma（MIA）from invasive adenocar⁃cinoma（IAC）presenting as pure ground⁃glass nodules（pGGNs）using CT imaging features and a neural networkmodel. Methods We retrospectively collected 151 patients with single MIA and 127 patients with single IACpresenting as pGGN on thin⁃slice CT surgically and pathologically confirmed in our hospital from January 2015 toJanuary 2018. The clinical and imaging features of all patients were collected and the differences of these featureswere compared between MIA group and IAC group. The input variables and factors based on p value less than 0.05with univariable analysis were used to establish a neural network model. The subject operating characteristic（ROC）curve was used to analyze，calculate and compare for predicting the diagnostic efficiency of IAC. Results Clinically，the average age of patients in MIA group was lower than that in IAC group（P < 0.001）. In imagingmorphology，the incidence of circular in MIA group was higher，but the incidences of lobulated sign，vesselconvergence sign，air bronchogram，cavitation and pleural indentation sign were less than in IAC group（all P <0.001）. In imaging quantitative analysis，mCT⁃LP，mCT⁃W，diameter，volume and mass were significantly lowerin MIA group than in IAC group（all P < 0.001）. The ROC curve of the neural network model was significantlybetter than the quantitative variables，with an AUC of 0.91，accuracy，sensitivity and specificity of 81.00% ，78.81% and 86.61%，respectively. Conclusions It is helpful to differentiate MIA from IAC presenting as pGGNsusing the neural network model.

Key words: ground?glass nodule, lung adenocarcinoma, CT, neural network