摘要Objective:To develop and validate a deep learning model for predict-ing the invasiveness of ground-glass opacity(GGO)lung adenocarcinoma based on spectral CT radiomic features.Methods:A total of 100 patients with pathologically confirmed GGO lung adenocarcinoma were retrospectively enrolled in this study,in-cluding 50 patients from our hospital(January 2020 to January 2023)and 50 external patients from collaborating hospitals.The in-house cohort was randomly divided into a training cohort(n=40)and an internal validation cohort(n=10)at a 4∶1 ratio,while the external 50 cases constituted the external validation cohort.Clinical baseline data and spectral CT images of all patients were collected.Two experienced radiologists independently segmented regions of interest(ROIs)using 3D Slicer software.Radiom-ic features were extracted via the Pyradiomics package,and features with an intraclass correlation coefficien(ICC)>0.8 were retained for subsequent analysis.The least abso-lute shrinkage and selection operator(LASSO)regression was performed to screen op-timal predictive radiomic features.Deep learning prediction models were constructed using deep neural networks based on the radiomic score(Rads)and clinical score(Clin).Receiver operating characteristic(ROC)curve analysis and decision curve anal-ysis(DCA)were applied to evaluate the predictive performance,discrimination ability,and clinical utility of the models.Results:A total of 19 stable spectral CT radiomic features were preliminarily screened via LASSO regression,among which 13 core features were finally adopted to establish the Rads through multivariate regression analysis.Univariate and multivariate logistic regression analyses identified age,maxi-mum tumor diameter,and lobulation sign as independent risk factors for high inva-siveness of GGO lung adenocarcinoma,which were used to construct the Clin model.Three predictive models(Rads model,Clin model,combined Rads+Clin model)were successfully established.ROC analysis demonstrated that all three models achieved an area under the curve(AUC)greater than 0.9 with excellent discriminative power.DCA results confirmed that model-guided clinical interventions yielded higher stan-dardized net benefits than universal intervention or non-intervention strategies.Fur-ther internal and external validation verified that the combined Rads+Clin model maintained AUC values above 0.9 across the training,internal validation,and exter-nal validation cohorts,without significant inter-cohort differences(P>0.05).Conclu-sion:The novel deep learning model integrating spectral CT radiomic features and clinical risk factors enables accurate and non-invasive preoperative prediction of in-vasiveness in GGO lung adenocarcinoma,which provides a reliable reference for in-dividualized clinical treatment decision-making and prognostic optimization.Future multi-center prospective studies are required to further improve the predictive accu-racy and generalizability of the model.