Abstract
OBJECTIVE: This study seeks to create and assess a combined radiomics model that combines intratumoral habitat features with peritumoral characteristics from CT imaging to predict spread through air spaces (STAS) in ≤ 2 cm solid lung adenocarcinomas. MATERIALS AND METHODS: A total of 401 patients with solid invasive lung adenocarcinomas ≤ 2 cm from two centers were retrospectively enrolled (training cohort: 217 cases, validation cohort: 93 cases, test cohort: 91 cases). Univariate and multivariate logistic regression analyses were employed to assess both CT features and clinical data, aiming to determine independent predictors of STAS. Regions of interest (ROI) for tumors were delineated on CT images, with peritumoral regions expanded by 1 mm, 3 mm, and 5 mm. Tumors were further segmented into three habitat subregions using K-means clustering. Radiomic features were extracted from the intratumoral, peritumoral, and habitat regions, and five machine learning algorithms were applied to construct predictive models. The best-performing predictive model was selected and further integrated into a combined model. Performance was assessed by receiver operating characteristic (ROC) curve's area under the curve (AUC), calibration curves, and decision curve analysis (DCA). RESULTS: The habitat model outperformed the Intra model, and the Peri3mm model surpassed Peri1mm and Peri5mm models. The integration of habitat, Peri3mm, and clinical models yielded a substantial improvement in predictive performance, with AUCs reaching 0.948, 0.897, and 0.930 in the training, validation, and test sets, respectively. Calibration curves and DCA confirmed favorable fit and higher clinical net benefit. CONCLUSION: The combined model provides high accuracy for predicting STAS in solid lung adenocarcinomas with a diameter of ≤ 2 cm, offering valuable support for treatment decision-making.