CT-based cross-center risk stratification of pure ground-glass nodules incorporating domain style-aware correction

BackgroundPure ground-glass nodules (pGGNs) are increasingly detected on chest CT. Although most exhibit indolent behavior, a subset may harbor invasive features before obvious radiological progression, making early risk stratification clinically important.PurposeTo develop and externally validate a CT-based deep learning model with domain adaptation for cross-center risk stratification of pGGNs.Material and MethodsThis retrospective multicenter study included 1235 surgically confirmed pGGNs from two institutions. A deep learning model incorporating domain style-aware correction was developed to differentiate AAH/AIS-like lesions from MIA/IAC-like lesions using preoperative CT images. Data from one center were used for training and internal validation, and an independent external cohort was used for validation. Model performance was evaluated using AUC, accuracy, calibration, and decision curve analysis.ResultsThe model achieved an AUC of 0.93 and an accuracy of 90.4% in the internal validation set. In the external validation set, the AUC was 0.88 with an accuracy of 81.7%, demonstrating good cross-center generalization. Domain adaptation improved external AUC from 0.85 to 0.88. Calibration analysis showed good agreement between predicted and observed probabilities. Decision curve analysis indicated higher net benefit compared with treat-all and treat-none strategies across a range of thresholds.ConclusionThe proposed model demonstrates robust performance across centers and may serve as a noninvasive tool to support individualized surveillance strategies for patients with pGGNs.