Spread through air spaces (STAS), a new invasive pattern in lung adenocarcinoma (LUAD), is a risk factor for poor outcome in early-stage LUAD. This study aimed to develop and validate a CT-based radiomics model for predicting STAS in stage IA LUAD.
A total of 395 patients (169 STAS positive and 226 STAS negative cases, including 316 and 79 patients in the training and test sets, respectively) with stage IA LUAD before surgery were retrospectively included. On all CT images, tumor size, types of nodules (solid, mix ground-glass opacities [mGGO] and pure GGO [pGGO]), and GGO percentage were recorded. Region of interest (ROI) segmentation was performed semi-automatically, and 1,037 radiomics features were extracted from every segmented lesion. Intraclass correlation coefficients (ICCs), Pearson’s correlation analysis and least absolute shrinkage and selection operator (LASSO) penalized logistic regression were used to filter unstable (ICC < 0.75) and redundant features (r > 0.8). A temporary model was established by multivariable logistic regression (LR) analysis based on selected radiomics features. Then, seven radiomics features contributing the most were selected for establishing the radiomics model. We then built two predictive models (clinical-CT model and MixModel) based on clinical and CT features only, and the combination of clinical-CT and Rad-score, respectively. The performances of these three models were assessed.
The radiomics model achieved good performance with an area under of curve (AUC) of 0.812 in the training set, versus 0.850 in the test set. Furthermore, compared with the clinical-CT model, both radiomics model and MixModel showed higher AUC and better net benefit to patients in the training and test cohorts.
The CT-based radiomics model showed satisfying diagnostic performance in early-stage LUAD for preoperatively predicting STAS, with superiority over the clinical-CT model.