Background Preoperative differentiation of pulmonary nodules into preinvasive (AAH/AIS), minimally invasive (MIA) and invasive adenocarcinoma (IAC) subtypes is vital for clinical decision-making, but conventional radiomics lacks biological interpretability and reproducibility. This study proposed a Moran's I-driven habitat radiomics approach to evaluate the temporal stability of invasiveness risk assessment on serial CT. Methods A total of 614 patients from two centers were enrolled, including a training set (n = 400), an internal validation set (n = 104), and an independent external testing set (n = 110). Notably, the external set comprised patients with serial longitudinal CT scans (preoperative, 3-, 6-, and 12-month) to validate temporal generalizability. Local Moran's I partitioned tumors into four habitats by spatial autocorrelation. Feature reproducibility was verified via image perturbation, and an optimal combined model was built with robust habitat features and compared with conventional radiomics. SHapley Additive exPlanation (SHAP) analysis was employed to revealed associations between habitat features and pathological cell density, offering hypothesis-generating biological insights. Results The Combined model demonstrated superior discrimination, achieving a macro-averaged AUC of 0.830 in the validation set and 0.854 in the external testing set. Specifically, the AUCs were 0.845 for AAH/AIS, 0.787 for MIA, and 0.931 for IAC. In the temporal robustness analysis, the Combined model outperformed conventional radiomics across all preoperative follow-up time points, yielding more reliable sensitivity for invasive lesions. SHAP analysis revealed that habitat features correlated with pathological cell density, offering intelligible biological insights. Conclusions This biologically plausible and temporally stable model enables noninvasive risk stratification of lung adenocarcinoma invasiveness and may support longitudinal surveillance of pulmonary nodules.
Gao et al. (Tue,) studied this question.