Unenhanced CT radiomics within a contrast-defined enhancing phenotype for differentiating noncalcified pulmonary hamartoma from peripheral solid invasive lung adenocarcinoma: a retrospective development and internal test study

Enhancing noncalcified pulmonary hamartoma (eNCPH) can mimic peripheral solid invasive lung adenocarcinoma (IAC). We assessed whether unenhanced CT radiomics could distinguish these lesions within a contrast-defined enhancing phenotype and explored a stacked model incorporating pleural indentation. This retrospective single-center study included 95 pathologically confirmed nodules (42 eNCPHs and 53 IACs) identified from May 2020 to January 2026 and assigned by a patient-level stratified random 7:3 split to development ( n = 66) and second-look internal evaluation ( n = 29). The 29 cases had been evaluated in an earlier analysis; no finalized-model parameter was optimized using them. After a reproducibility gate and training-only feature reduction, five classifier families were compared, and an eight-feature logistic regression (LR) model was assessed by repeated 5-fold × 10 nested cross-validation. A stacked model combined pleural indentation with repeated cross-fitted LR probabilities. Frozen models were applied to the 29 cases. LR ranked first in full-development model-selection cross-validation (mean AUC, 0.786; SD, 0.142), although classifier-family selection varied across repeated nested splits. In the second-look internal evaluation, AUCs were 0.697 (95% CI, 0.526–0.868) for the morphologic model, 0.769 (95% CI, 0.583–0.955) for standalone radiomics, and 0.827 (95% CI, 0.665–0.989) for the stacked model. Pairwise DeLong comparisons were not significant after Holm adjustment (all adjusted P ≥ 0.150). The combined model achieved 87.5% sensitivity, 69.2% specificity, and 79.3% accuracy. The standalone radiomics calibration slope was 0.026 (95% CI, − 0.105 to 0.157), indicating severe probability miscalibration despite moderate discrimination; its probabilities should not be interpreted as absolute diagnostic probabilities. Within this selected contrast-defined phenotype, unenhanced CT radiomics showed moderate discrimination. The stacked model had a numerically higher AUC, but incremental value was not established. The small second-look evaluation, severe standalone-model miscalibration, feature-selection instability, and acquisition heterogeneity require external validation before clinical use.