Deep Learning Model with Nodule Indexing Tailored to Early-Stage Lung Cancer Detection.

Authors

Schroeder JL, Cormier MG, Lo SB, Gillis LB, Freedman MT, Mun SK

Affiliations

• Radiology Department, Georgetown University Medical Center, Washington, DC 20007. Electronic address: JamieLee.T.Schroeder@medstar.net.
• Radiology Department, Georgetown University Medical Center, Washington, DC 20007. Electronic address: mcormi01@gunet.georgetown.edu.
• Arlington Innovation Center: Health Research, Virginia Polytechnic Institute and State University - D.C. Area, Arlington, VA 22203; Oncology Department, Lombardi Comprehensive Cancer Center, Georgetown University Medical Center, Washington, DC 20007.
• NXGEN BIOSTATS, LLC, Mount Airy, Maryland 21771. Electronic address: laura.gillis.nxgenbiostats@gmail.com.
• Radiology Department, Georgetown University Medical Center, Washington, DC 20007; Oncology Department, Lombardi Comprehensive Cancer Center, Georgetown University Medical Center, Washington, DC 20007. Electronic address: mtfreedman@verizon.net.
• Arlington Innovation Center: Health Research, Virginia Polytechnic Institute and State University - D.C. Area, Arlington, VA 22203. Electronic address: munsk@vt.edu.

Abstract

OBJECTIVE: To evaluate whether a deep learning-based AI system with suspected nodule indexing and malignancy risk stratification improves radiologist performance in detecting pulmonary nodules on CT, using a dataset enriched with challenging early-stage lung cancers.

METHODS: The study comprised a standalone AI sensitivity-specificity analysis and a two-arm crossover reader study with 16 American board-certified radiologists. Each reader interpreted 340 CT scans with and without AI, separated by a one-month washout. The dataset included 209 screening and 131 non-screening cases: 133 with lung cancer, 61 with benign non-calcified nodules ≥4 mm, and 146 normal. To enrich subtle lesions, 64 of 91 (70.3%) small cancer cases were drawn from early-round NLST CT scans. Localization-specific ROC (LROC) analysis was used to assess radiologist performance.

RESULTS: Standalone AI achieved a sensitivity of 0.804 at 1.37 false positives per case. With AI assistance, radiologists' LROC AUC improved cancer detection (0.761 vs. 0.652; ΔAUC = 0.109, 95% CI: 0.067, 0.152) and for all nodules (0.830 vs. 0.734; ΔAUC = 0.096, 95% CI: 0.059, 0.133). Mean sensitivity increased from 0.585 to 0.727, while specificity remained essentially unchanged (0.918 vs. 0.913). Interpretation time decreased by 12.9%, from a mean of 133 to 115.9 seconds (Difference = -17.1 seconds (95% CI: -26.7, -9.0)). AI alerts enabled detection of early-stage cancer detection previously missed in NLST interpretations.

DISCUSSION: The AI system significantly improved radiologist's performance in pulmonary nodule detection, with consistent benefits across nodule types, screening contexts, and experience levels; supporting its integration into routine chest CT interpretation workflows.