Machine learning in differentiating pulmonary invasive adenocarcinoma from non-invasive adenocarcinoma manifested as pure ground-glass nodule / 中国医学影像技术

ABSTRACT

Objective: To investigate the value of machine learning model based on radiomic features in differentiating pulmonary invasive adenocarcinoma from non-invasive adenocarcinoma manifested as pure ground-glass nodule (pGGN). Methods: A total of 87 lung adenocarcinomas CT presented as pGGN were analyzed retrospectively, including 32 cases with invasive adenocarcinoma (IAC) and 55 cases with non-IAC (17 adenocarcinomas in-situ [AIS] and 38 minimally invasive adenocarcinomas [MIA]). The software ITK-SNAP was used to draw ROI, and the radiomic features were extracted using A.K. analysis software. After screening the significant characteristic parameters, the feature dimensionality reduction was conducted with Spearman analysis and Lasso regression. The final feature parameters were selected to construct three machine learning models, including support vector machine (SVM), random forest (RF) and logistics regression (LR). Then 10-fold cross validation was used to get the optimal model, and ROC curve was drawn to evaluate the performance of 3 models. Results: A total of 396 radiomic features were extracted, and 19 features were finally obtained after feature screening. Machine learning models SVM, RF and LR could effectively distinguish IAC from non-IAC, with the accuracies of 93.30%, 86.70% and 83.30%, and AUC of 0.94, 0.92 and 0.83 respectively. Conclusion: Machine learning models based on radiomic features have good classified performances, which can effectively distinguish IAC from non-IAC manifested as pGGN preoperation.