Computed tomography-based tumoral and peritumoral radiomics models for preoperative prediction of the spread through air spaces in patients with clinical stage I pure solid invasive lung adenocarcinoma: a multicenter study.

Background: Spread through air spaces (STAS) is recognized as a novel invasive mode of lung adenocarcinoma (LADC), linked with poorer prognosis and high risk of recurrence. The aim of this study was to develop and evaluate a radiomics nomogram using computed tomography (CT)-based tumoral and peritumoral radiomics features for preoperatively predicting STAS status in clinical stage I pure-solid LADC. Methods: This study retrospectively enrolled 308 individuals with stage I LADC appearing as pure-solid nodules on thin-section CT who underwent surgical resection from three institutions. We randomly split the patients at authors' hospital into a training set (n=174) and internal validation set (n=73) in a ratio of 7:3, while the external validation set consisted of 61 patients from the other two hospitals. The radiomics features extracted from the gross tumor volume (GTV), two types of peritumoral tumor volume (PTV) (5 and 10 mm around the tumor), and their corresponding two types of gross peritumoral tumor volume (GPTV) were utilized to construct five radiomics models, respectively. Univariate and multivariate analyses identified the independent predictors of STAS. The radscore of the radiomics model with optimal performance was integrated with clinical predictor to develop a comprehensive nomogram. Results: The STAS positive status was found in 118 (38.3%) of the 308 patients {female: 54.2%; median [interquartile range (IQR)] age: 65, [57-72] years}. The GPTV10 model achieved the highest area under the curve (AUC) values of 0.741, 0.737 and 0.741 in three cohorts. The multivariate logistic regression (LR) suggested that micropapillary component was the independent risk factor of pathological STAS. The comprehensive model constructed using the GPTV10 radscore and clinical predictor exhibited AUCs of 0.788, 0.748 and 0.783. The decision curve analysis (DCA) revealed that the nomogram had superior capacity for predicting STAS status in LADC. Furthermore, both pathological STAS status and STAS predicted by the combined model stratified patients for prognosis, with 5-year recurrence-free survival (RFS) showing obvious difference between STAS-positive and STAS-negative. Conclusions: Peritumoral features were significantly correlated with STAS status. The integration of radiomics characteristics and clinical factor provided better performance in the prediction of STAS status.