A nomogram model for preoperative evaluation of invasive degree in lung adenocarcinoma: integrating intratumoral and 2 mm peritumoral radiomics scores with conventional features

Objective To determine the optimal peritumoral margin for predicting invasiveness in lung adenocarcinoma and to develop and validate a nomogram integrating conventional CT features with intratumoral and peritumoral radiomics scores for non-invasive preoperative assessment of invasive degree. Methods This retrospective study included 171 patients with pathologically confirmed pulmonary nodules from the Affiliated Hospital of Liaoning University of Traditional Chinese Medicine (January 2018 – December 2025), randomly split into training ( n = 120) and test ( n = 51) sets (7:3 ratio). Based on the 2021 WHO classification, nodules were classified as pre-microinvasive (AAH, AIS, MIA) or invasive (IAC). Conventional CT features and 107 radiomics features were extracted from intratumoral and varying peritumoral margins (0–2, 0–4, 0–6, and 0–8 mm). Feature selection was performed using LASSO and logistic regression to construct five models: a conventional model, an intratumoral radiomics score model, peritumoral radiomics score models (at different margins), an intratumoral-peritumoral radiomics score model, and a combined conventional-intratumoral-peritumoral radiomics model. Performance was assessed via ROC curves, DCA, calibration curves, nested 5-fold cross-validation (25 iterations) and nested Bootstrap (1,000 iterations). Model interpretability was analyzed using SHAP (SHapley Additive exPlanations). Results Multivariate logistic regression identified maximum diameter and mean CT value as independent predictors. Among the four peritumoral ranges, the 2 mm peritumoral radiomics score model showed optimal performance (AUC: 0.902 training, 0.881 test), establishing 2 mm as the optimal margin. The combined model (conventional + intratumoral + 2 mm peritumoral) achieved the best predictive performance (AUC: 0.925 training, 0.888 test), outperforming the conventional and single radiomics models. The model demonstrated good calibration (Hosmer-Lemeshow test, P > 0.05). To prevent information leakage, nested 5-fold cross-validation (25 iterations) and nested Bootstrap (1,000 iterations) were performed on the training set, yielding a mean AUC of 0.873 ± 0.079 and 0.763 (95% CI: 0.572–0.907), respectively. DCA confirmed a favorable net clinical benefit. SHAP analysis identified the intratumoral-2 mm peritumoral radiomics score, mean CT value, and maximum diameter as the top three contributing features. Conclusion A 2 mm peritumoral margin is optimal for predicting lung adenocarcinoma invasiveness. The proposed nomogram enables non-invasive, accurate preoperative prediction with strong interpretability and clinical utility, serving as a reliable tool for guiding individualized surgical planning.