1266P Prediction of Lung Cancer Risk in Ground-Glass Nodules Using Deep Learning from CT Images

Pulmonary nodules (PNs) are frequently detected on CT scans and may indicate early lung cancer. Among PNs, management of ground-glass nodules (GGNs) can be particularly challenging due to their long, often indolent natural course. When malignant, they are usually lepidic adenocarcinomas, for which many targeted therapies exist. Artificial Intelligence (AI)-based malignancy prediction of PNs using CT images has been demonstrated to perform well on solid and part-solid PNs. Here, we aimed to extend one such AI model to GGNs. GGN nodules and predominantly-GGN part-solid nodules were collected from 3 sites: two UK, one US. Patients with a 5-30mm nodule on non-contrast CT were included, unless they had a history of cancer in the past 5 years, >5 nodules, or insufficient outcome data from long-term follow-up or histological diagnosis. Scans with >2.5mm slice spacing or significant imaging artifacts were excluded. The most concerning nodule on each CT study was evaluated, giving 169 cancers in 57 patients, and 347 benign nodules in 180 patients. Most patients were enrolled through a text search of radiology reports, and enriched for malignant nodules based on local data. The AI tool is based on a commercial system, further trained using GGNs from an additional UK site. Ability to predict malignancy was evaluated using the Area Under the Receiver Operating Characteristic curve (AUC) and compared to two reference models. The ability to rule benign cases out of follow-up was evaluated by calculating specificity at 100% sensitivity. The AUC of the AI model was 89.1 [95%CI, 86.3, 91.8], while the AUC of the reference models, Brock and Mayo, were 86.5 95%CI [95% CI, 83.2, 89.6; p=0.39 vs. AI] and 80.9 [95%CI76.9, 84.8; p<0.01 vs. AI] respectively. The percentage of benign GGNs ruled out for cancer at 100% sensitivity were 48.7% for the AI model, 14.1% for Brock, and 16.7% for Mayo. To our knowledge, this is the first study to demonstrate the potential value of an AI model in predicting the malignancy risk of GGNs. Implementation of such models in clinical practice could lead to more appropriate use of invasive interventions among GGN patients, thereby improving patient outcomes.