Comparison of Two Deep Learning Models and Their Combination for 1-year Lung Cancer Detection in Two Screening Cohorts.
Authors
Affiliations (12)
Affiliations (12)
- Department of Electrical and Computer Engineering, Cornell University, Ithaca, NY.
- HeartLung.AI, 2450 Holcombe Blvd, Houston, TX 77021.
- Tustin Teleradiology, Tustin, Calif.
- Department of Diagnostic, Molecular, and Interventional Radiology, Icahn School of Medicine at Mount Sinai, New York, NY.
- Department of Radiology, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands.
- Pulmonary Department, Clínica Universidad de Navarra, Pamplona, Spain.
- CIBER for Respiratory Diseases (CIBERES), Carlos III Health Institute, Madrid, Spain.
- Department of Biomedical and Molecular Sciences, Queen's University, Kingston, ON, Canada.
- Division of Respirology, Department of Medicine, Queen's University, Kingston, ON, Canada.
- Kingston Health Sciences Centre, Kingston, ON, Canada.
- Translational Research in Respiratory Medicine, IRBLleida, Arnau de Vilanova and Santa Maria University Hospital, Lleida, CIBER for Respiratory Diseases (CIBERES), Carlos III Health Institute, Madrid, Spain.
- Clínica Universidad de Navarra (CUN), Madrid, Spain.
Abstract
Purpose To evaluate the performance of Sybil, a Medical Open Network for Artificial Intelligence (MONAI)-based nodule model, and their combination for 1-year lung cancer detection in two screening cohorts: the National Lung Screening Trial (NLST) and the Pamplona International Early Lung Cancer Action Program (P-IELCAP). Materials and Methods Baseline low-dose CT (LDCT) examinations from the NLST and the P-IELCAP were analyzed. A prespecified OR-stacked composite model combining Sybil, a MONAI-based nodule model, and a safety-net rule (MONAI-detected nodule > 6 mm with detection confidence > 0.95) was developed and evaluated for 1-year lung cancer detection. Model discrimination was assessed by using area under the receiver operating characteristic curve (AUC), and operating-point performance metrics were calculated. Results A total of 5110 baseline LDCT examinations were analyzed, including 2343 from NLST and 2767 from P-IELCAP, with 65 lung cancers diagnosed within 1 year. Mean age, mean pack-years, and percentage of female participants were 61.6 years ± 5.1 (SD), 57.3 pack-years ± 24.8, and 39.4% (923 of 2343), respectively, in NLST and 56.0 years ± 9.5, 34.4 pack-years ± 21.5, and 28.6% (791 of 2767), respectively, in P-IELCAP. Lung cancer incidence (per 1000 person-years) was higher in NLST than in P-IELCAP (17.7 vs 8.7; <i>P</i> = .006). Mean Sybil 1-year lung cancer probabilities (scaled by 1000 for presentation) were also higher in NLST than in P-IELCAP (16.4 vs 9.9; <i>P</i> < .01). Sybil achieved an AUC of 0.93 in both cohorts, whereas the MONAI-based model achieved AUCs of 0.90 in NLST and 0.79 in P-IELCAP. The composite model missed two of 65 1-year lung cancers while correctly classifying 2731 of 4965 (55.0%) noncancer examinations as negative, yielding a sensitivity of 63 of 65 (96.9%) and a positive predictive value of 63 of 2297 (2.7%). Conclusion For 1-year lung cancer detection with LDCT, the composite AI model achieved high sensitivity while maintaining acceptable specificity. <b>Keywords:</b> Applications-CT, Thorax, Pulmonary, Sybil, MONAI, NLST, P-IELCAP, Lung Cancer, Low Dose Computed Tomography <i>Supplemental material is available for this article.</i> © RSNA, 2026.