Machine learning models based on tumoral, 4 mm-peritumoral, and combined radiomic features for evaluating the invasiveness of lung adenocarcinoma manifesting as ground-glass nodules.

backgroundLung minimally invasive adenocarcinoma (MIA) and invasive adenocarcinoma (IAC) manifest as ground-glass nodules (GGNs) on computed tomography scans, but their invasiveness, treatment modalities, and prognosis are different. This study used machine learning approaches to construct models based on tumoral or 4 mm-peritumoral radiomic features of GGNs or their combination and assessed their value in evaluating the invasiveness of lung adenocarcinoma.

methodsIn total, 287 patients with GGNs confirmed as MIA or IAC were retrospectively included and randomly classified into training and test sets at a 7:3 ratio. Radiomics features of GGN were extracted from the tumoral and 4-mm peritumoral regions. Eight machine learning approaches (logistic regression, random forest, support vector machine, adaptive boosting, k-nearest neighbor, decision tree, naive Bayes, and neural network) were used to construct models based on tumoral, 4 mm-peritumoral, and combined radiomic features.

resultsEleven tumoral and eight 4 mm-peritumoral radiomic features were selected. In the training set, models constructed with tumoral [area under the curve (AUC): 0.895–0.990], 4 mm-peritumoral (AUC: 0.909–0.985), and combined radiomic features (AUC: 0.895–0.987) showed excellent values to differentiate MIA from IAC. In the test set, these models also had excellent ability in distinguishing MIA from IAC whether they were constructed with tumoral (AUCs for LR, RF, SVM, AB, KNN, DT, NB, and NN were 0.904, 0.854, 0.886, 0.870, 0.884, 0.802, 0.754, and 0.903), 4 mm-peritumoral (AUCs for LR, RF, SVM, AB, KNN, DT, NB, and NN were 0.854, 0.843, 0.836, 0.862, 0.820, 0.791, 0.842, and 0.855) or combined radiomic features (AUCs for LR, RF, SVM, AB, KNN, DT, NB, and NN were 0.913, 0.873, 0.893, 0.875, 0.893 0.803, 0.801, and 0.902). Most machine learning models showed no difference in AUC in pairwise comparisons via the DeLong test.

conclusionsMachine learning models constructed with tumoral, 4 mm-peritumoral, and combined radiomic features of GGNs are valuable for distinguishing MIA from IAC.