Multicenter validation of a CT-based deep learning model to inform treatment decision-making for pure ground-glass pulmonary nodules

Background: Distinguishing invasive adenocarcinoma (IAC) from non-IAC in pure ground-glass nodules (pGGNs) remains a critical clinical challenge. We aimed to develop and multicenter-validate a CT-based deep learning model to differentiate IAC from non-IAC lesions in pGGNs and compare its performance with that of human experts, thereby identifying candidates for definitive surgical management among pGGNs likely to represent IAC. Methods: This retrospective study included 1707 surgically resected pathologically confirmed pGGNs from six medical institutions. We developed Lung-PNetV2, a modular deep-learning framework that integrates cross-scanner normalization, 3D volumetric encoding (via ResNet-18), and the multimodal fusion of imaging, nodular, and clinical features. The model was trained on 847 pGGNs and validated internally (203 pGGNs) and externally (657 pGGNs). Seven clinicians (four radiologists and surgeons) independently evaluated the holdout test set using the NCCN-guided 5-point scoring. Results: Lung-PNetV2 achieved AUCs of 0.892 (training), 0.831 (internal test), and 0.827 (external test), significantly outperforming all the human readers (AUC: 0.681–0.722; P < 0.01). At the clinical decision threshold (0.6 probability for IAC), the model demonstrated balanced performance in the external test set: 81.0% accuracy, 67.7% sensitivity, 84.1% specificity, and 91.8% negative predictive value. The deep learning model surpassed radiologists in sensitivity (67.7% vs 60.9%) and surgeons in specificity (84.1% vs 78.3%) while maintaining superior F1-macro (72.5% vs reader range: 54.1–67.0%; P < 0.05, 4/7 readers). Conclusion: CT-based Lung-PNetV2, a deep learning model, provides a generalizable performance that surpasses that of human experts in stratifying the invasiveness of pGGNs through the effective cross-modal fusion of imaging and clinical data. Its balanced sensitivity and specificity profiles support risk-stratified management, allowing predicted non-IAC pGGNs to avoid unnecessary procedures, whereas high-risk cases receive timely definitive surgery, enabling individualized and precise treatment.