463P - Prediction of invasiveness in lung adenocarcinoma using machine learning algorithm based on 3D-CT imaging

Abstract

BackgroundT factor depends on consolidation diameter in lung adenocarcinoma. Adenocarcinoma with short consolidation diameter could be indicated for sublobar resection. However, measurement of consolidation diameter under lung window in 2D-CT could be problematic, as the outline is not clear. We evaluated the relationship between consolidation lesion under mediastinal window in 3D-CT and pathological invasiveness, and aimed to predict invasiveness using machine learning algorithm. MethodsNinety-five patients who underwent surgical resection of lung adenocarcinoma that is less than 20 mm in diameter and has consolidation were analyzed retrospectively, in our Hospital from 2010 to 2016. The total tumor diameter and the volume and diameter of consolidation were analyzed using thin-slice CT and 3D-CT. Preoperative FDG-PET and serum CEA were also measured. Preinvasive lesion and minimally invasive adenocarcinoma were classified as pathologically non-invasive group, and invasive adenocarcinoma was classified as pathologically invasive group. The statistical differences were assessed by the Mann-Whitney U test or chi-square test. We used two machine learning algorithm. Three logistic regression (LR) models were trained based on sex, total tumor diameter, SUVmax, CEA, and (1) consolidation diameter (LR-Cd), (2) consolidation volume (LR-Cv), (3) consolidation diameter and volume (LD-Cd+Cv). Similarly, three random forest (RF) models were trained (RF-Cd, RF-Cv, RF-Cd+Cv). Patients were split into a training set (n = 76) and test set (n = 19). Model performance was measured area under the curve (AUC), and compared using receiver-operating characteristics curve analysis on the test set. ResultsTwenty-six non-invasive adenocarcinomas and 69 invasive adenocarcinomas were evaluated. The consolidation diameter and volume, and SUVmax of invasive group were significantly greater than those of non-invasive group (p < 0.001). On the test set, AUC of LR-Cd, RF-Cd, LR-Cv, RF-Cv, LR-Cd+Cv and RF-Cd+Cv were 0.687, 0.703,0.628, 0.744, 0.628 and 0.744 respectively. ConclusionsThe pathological invasiveness could be predicted by using machine learning algorithm. Especially, higher AUCs were obtained in RF-Cv and RF-Cd+Cv model. Legal entity responsible for the studyThe authors. FundingHas not received any funding. DisclosureAll authors have declared no conflicts of interest.