Abstract: Objective To develop a prediction model based on clinical features and preoperative computed tomography(CT), which would serve as a non-invasive method for diagnosing lung cancers associated with cystic airspaces(LCCA). Methods Clinical data of patients undergoing surgery during Jan. 2020 and Dec. 2021 were retrospectively reviewed. A total of 296 patients were enrolled and divided into LCCA group and benign group. Radiomics features were extracted from lesions on preoperative CT images, and clinicopathological characteristics were carefully analyzed. The aforementioned features were used to construct a diagnostic model with random forest algorithm following feature reduction. The 214 patients from the Affiliated Hospital of Qingdao University underwent training and internal verification(D1). The remaining 82 patients from the Affiliated Hospital of Jining Medical College served as an independent external validation cohort(D2). The classification ability of the model was evaluated with receiver operating characteristic(ROC)curve, calibration curve and decision curve analysis(DCA). Results There were no other significant differences between the two groups in clinical features except that the LCCA group averaged older than the benign group(60.12±9.95 vs. 50.86±13.66, P<0.001). After radiomic features were screened, the following features were obtained, including shape flatness, robust mean absolute deviation and difference variance. The model demonstrated superior performance in discriminating LCCA in both D1(AUC 0.97)and D2(AUC 0.74), with accuracy, sensitivity and specificity on D1 being 0.92, 0.85 and 0.99, respectively. Conclusion We have developed and externally validated a practical model that can accurately differentiate LCCA from benign lesions. This innovative approach may serve as a new non-invasive diagnostic tool for LCCA.

Key words: Lung cancer associated with cystic airspaces, Radiomics, External validation, Computed tomography