PET/CT radiomics for noninvasive histologic risk stratification in resected lung adenocarcinoma: a single-system development and internal validation study
Year: 2026
Type: article
Study design: Observational Study
Abstract: Background: F-FDG PET/CT) radiomics could predict histologic patterns and pathologic risk features using a hierarchical modelling strategy. Methods: This retrospective cohort study included patients from three imaging facilities within a single academic health system. Radiomic features were extracted from preoperative PET/CT and CT segmentations; preprocessing included feature filtering, standardisation, redundancy reduction, and ComBat harmonisation. Penalised generalised linear model (GLMNET), random forest, and XGBoost models were trained with stratified cross-validation and evaluated in independent test sets. Outcomes included mucinous phenotype, non-mucinous histologic subtype, histologic grade, spread through air spaces (STAS), vascular invasion, and visceral pleural invasion (VPI). Eligible patients had primary lung adenocarcinoma treated by curative-intent resection, with preoperative PET/CT or thin-section CT and histopathologic review available; patients receiving neoadjuvant therapy, non-adenocarcinoma histology, or rare adenocarcinoma variants were excluded. The histologic reference standard was central review by a single thoracic pathologist per the 2021 WHO/IASLC (World Health Organization/International Association for the Study of Lung Cancer) classification, blinded to imaging and radiomic results. From 300 candidate radiomic variables (147 PET-derived, 153 CT-derived), models were developed using a 75%/25% stratified train-test split within the same single-institution cohort; the test set therefore represents internal, not external, validation. Results: grade 1-2) achieved area under the receiver operating characteristic curve (ROC-AUC) 0.767 [95% confidence interval (CI): 0.624-0.897], area under the precision-recall curve (PR-AUC) 0.854 (95% CI: 0.723-0.952), and satisfactory calibration [Brier Score 0.170; Brier Skill Score (BSS) 0.23]. VPI yielded a moderate global signal (ROC-AUC 0.691, 95% CI: 0.518-0.841). Non-mucinous subtype classification showed preserved probabilistic ranking despite limited categorical accuracy (balanced accuracy 0.614). Models for mucinous phenotype, STAS, and vascular invasion showed limited discrimination; VPI in the small-tumour subgroup was constrained by insufficient positive events. Conclusions: PET/CT radiomics supported probabilistic preoperative risk stratification of lung adenocarcinoma, particularly for identifying poorly differentiated tumours. These results support radiomics as a complementary preoperative tool pending external validation.
Language: English
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: diamond
OpenAlex ID: W7224015167