Abstract
BACKGROUND: This study aims to explore the value of radiomic features from different regions of part-solid nodules (PSNs) for predicting spread through air spaces (STAS) in lung adenocarcinoma. METHODS: This retrospective analysis included 333 patients with PSNs lung adenocarcinoma pathologically confirmed in three hospitals. Data from one institution were utilized for training set (n=223), while the remaining two served as the external test set (n=110). The computed tomography radiomic features were extracted from different areas of the nodule (ground-glass, solid, gross, and perinodular). Three machine learning classifiers (support vector machine, light gradient boosting machine [LightGBM], logistic regression) were used to build predictive models. Model performance was assessed using accuracy and area under the curve (AUC). The DeLong test was used to determine differences in AUC values between models. The clinical benefits of models were assessed using decision curve analysis (DCA). RESULTS: In the external test set, the radiomics model developed using combined features from ground-glass, solid, and perinodular regions with LightGBM classifier achieved an AUC of 0.840 (95% confidence interval [CI]: 0.758-0.921), which was better than the clinical model (AUC = 0.622, 95% CI: 0.494-0.750, P < 0.001) and other radiomics models. DCA indicated that this model has achieved a higher net benefit. CONCLUSION: The radiomics model developed using radiomic features of distinct solid and ground-glass components of PSNs and the perinodular region can contribute to identifying the STAS status in lung adenocarcinoma.