LungAI: An AI Biomarker that Refines Lung-RADS v2022 Triage of Screen-Detected Pulmonary Nodules

Background: The Lung-RADS algorithm standardizes the management of screen-detected pulmonary nodules, but its risk assessment is driven chiefly by nodule size and attenuation. This misallocates a meaningful fraction of patients: a nontrivial share of small malignancies receive 12-month surveillance while large benign nodules proceed to invasive workup. Artificial intelligence (AI) may supply risk information independent of the imaging features Lung-RADS already uses. Methods: We validated LungAI, an investigational deep learning nodule malignancy classifier not yet cleared by the US Food and Drug Administration (FDA), on the held-out NLST test partition of Ardila et al. (2,747 low-dose CT examinations from 1,244 patients with at least 1 solid or part-solid 4-30 mm nodule; 89 pathologically confirmed cancers, prevalence 3.2%). We compared discrimination against Sybil, mean nodule diameter, and estimated nodule volume. A board-certified radiologist assigned Lung-RADS v2022 categories to a positives-enriched random subset (226 eligible examinations; 78 cancers), projected to the full cohort by contingency weighting. A per-category reclassification algorithm shifted patients along the existing Lung-RADS management ladder using a-priori score cutpoints. We evaluated a separately locked follow-up configuration on examinations with a prior scan. Results: LungAI achieved AUROC 0.92 (95% CI 0.88-0.95) and AUPRC 0.47 (0.35-0.61), outperforming Sybil (delta AUROC 0.07; delta AUPRC 0.18), diameter (delta AUROC 0.14; delta AUPRC 0.34), and volume (delta AUROC 0.14; delta AUPRC 0.38; all P<0.001). Against Lung-RADS v2022 scored ordinally on the weighted annotated subset, LungAI improved AUPRC (delta AUPRC 0.31, P=0.003) but not AUROC (delta AUROC 0.03, P=0.20). LungAI added discriminative information independent of diameter (likelihood-ratio chi squared = 161, P < 10e-36; delta AUROC 0.10) and, on follow-up examinations, independent of delta diameter, volume doubling time, and new-nodule status (chi squared = 79, P < 10e-18). Integrated with Lung-RADS, LungAI reduced cancers on the slowest (12-month rescan) pathway by 46% (3.1 --> 1.6 per 1,000 examinations) and benign biopsy referrals by 20% (32.7 --> 26.2 per 1,000 examinations) at approximately unchanged total biopsy volume (51.0 --> 50.1 per 1,000). Assuming full adherence to recommended follow-up, the combined approach yielded 25 months of net diagnostic lead time, an estimated net cost saving of ~$2,200, and 6.5 fewer benign biopsies per 1,000 examinations; accounting for published adherence rates, it yielded 1.6 additional cancer diagnoses within 1 year per 1,000 examinations. Conclusions: LungAI contributes malignancy-risk information orthogonal to the size- and growth-based features underlying Lung-RADS and improves the allocation of screening patients across the existing management spectrum while potentially decreasing invasive procedures for benign lesions. Multi-site, multi-geography validation is required to demonstrate practical generalizability.