3D intratumoral heterogeneity-based quantitative score from chest CT for preoperative prediction of visceral pleural invasion in lung adenocarcinoma: a multicenter study.

Introduction: Visceral pleural invasion (VPI) is a critical adverse prognostic factor in lung adenocarcinoma (LUAD). This study aimed to develop a stacking ensemble model that integrates three-dimensional intratumoral heterogeneity (3D ITH) scores with clinicoradiologic features to achieve accurate preoperative prediction of VPI in LUAD. Methods: This multicenter retrospective study included 1,301 patients with LUAD from three medical centers. Patients from Centers 1 and 2 were assigned to the development cohort, whereas those from Center 3 constituted the fixed external validation cohort. To calculate the 3D ITH score, we integrated local radiomic descriptors with global pixel distribution characteristics derived from whole tumor CT volumes. Clinicoradiologic features and 3D ITH scores were then used to construct six base machine learning models and a final stacking ensemble classifier. Model performance was primarily assessed using receiver operating characteristic analysis and the area under the curve (AUC). SHapley Additive exPlanations (SHAP) were used to quantify feature contributions and to interpret the final model. Results: The stacking ensemble classifier achieved the highest AUC for preoperative prediction of VPI in LUAD (AUC = 0.878), whereas XGBoost showed competitive performance on several threshold dependent metrics. SHAP analysis identified the 3D ITH score as the most influential predictor, followed by nodule size and CT density. Comparative experiments further showed that the stacking ensemble model outperformed the conventional radiomics signature (AUC = 0.841) and the clinicoradiologic comparative model (AUC = 0.776). Conclusion: The model integrating 3D ITH scores with clinicoradiologic features showed strong discrimination for preoperative prediction of VPI in LUAD. This approach may serve as a useful adjunct for preoperative risk stratification and individualized treatment planning.