CT radiomics for stratifying pulmonary adenocarcinoma-spectrum lesions presenting as ground-glass nodules: a single-center retrospective study

Preoperative estimation of invasiveness in ground-glass nodules (GGNs) remains challenging. We evaluated computed tomography (CT) radiomic models at three prespecified pathological thresholds across atypical adenomatous hyperplasia (AAH) and pulmonary adenocarcinoma-spectrum lesions. This retrospective single-center study included 253 patients with surgically resected GGNs: 19 AAH, 34 adenocarcinoma in situ (AIS), 47 minimally invasive adenocarcinoma (MIA), and 153 invasive adenocarcinoma (IAC). The cohort was randomly allocated 4:1 with stratification by pathological class into a training set ( n = 202) and a held-out internal test set ( n = 51); the training/test class counts were 15/4 AAH, 27/7 AIS, 38/9 MIA, and 122/31 IAC. Three binary pathological-threshold tasks were evaluated: AAH versus AIS/MIA/IAC (Task 1), AAH/AIS versus MIA/IAC (Task 2), and AAH/AIS/MIA versus IAC (Task 3). Three-dimensional regions of interest were manually delineated on thin-section CT. Radiomic features were reduced by zero-variance removal, univariate screening, Pearson correlation filtering, and least absolute shrinkage and selection operator (LASSO) regression, followed by random forest classification. In a secondary exploratory analysis, a fixed ImageNet-pretrained ResNet18 generated 512-dimensional transfer features that were selected by LASSO and entered into a random forest classifier. Performance was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, predictive values, and F1 score. Of 1,836 initial radiomic features per lesion, 10, 14, and 10 were selected for Tasks 1, 2, and 3, respectively. Test-set AUCs were 0.9016 (95% CI 0.7927–1.0000), 0.8898 (95% CI 0.7992–0.9803), and 0.9484 (95% CI 0.8888–1.0000). For Task 3, accuracy was 0.8824, sensitivity 0.8387, and specificity 0.9500. Pooled five-fold ResNet18 AUCs were 0.8520, 0.8203, and 0.8338 for Tasks 1, 2, and 3, respectively. In this single-center surgically selected cohort, CT radiomic models showed internal discrimination across three pathological-threshold tasks, including discrimination of IAC from AAH/AIS/MIA in Task 3. These results do not establish clinical utility and should be interpreted in light of the small AAH subgroup, class imbalance, single-center sampling, and lack of external validation. The ResNet18 analysis was exploratory because it used display-export images, non-nested feature selection, and a validation framework different from that of the primary radiomics analysis.