Comparison of Two Deep Learning Models and Their Combination for One-Year Lung Cancer Detection in Two Screening Cohorts

Submitted manuscript | Original research review

This study compares two complementary deep-learning approaches for detecting lung cancer within one year of baseline low-dose CT (LDCT) screening: Sybil, which analyzes the whole scan, and a MONAI-based model focused on pulmonary nodules. It also evaluates a prespecified composite model designed to combine their signals.

Study design

The retrospective analysis included 5,110 baseline LDCT examinations from two screening cohorts: the US National Lung Screening Trial (NLST; 2,343 examinations) and the Pamplona International Early Lung Cancer Action Program (P-IELCAP; 2,767 examinations). The composite OR-stacked model combined Sybil, the MONAI nodule model, and a safety-net rule for MONAI-detected nodules larger than 6 mm with confidence above 0.95. The outcome was histologically confirmed lung cancer within one year.

Key findings

Sixty-five lung cancers were diagnosed within one year. Sybil showed an AUC of 0.93 in each cohort; the MONAI-based model achieved AUCs of 0.90 in NLST and 0.79 in P-IELCAP. The composite model detected 63 of 65 cancers (sensitivity 96.9%) and classified 2,731 of 4,965 non-cancer examinations as negative (specificity 55.0%). Its positive predictive value was 2.7%.

Interpretation

Combining whole-scan and nodule-centered AI yielded high sensitivity across these two cohorts, with moderate specificity. Because this was a retrospective evaluation in existing screening datasets, prospective assessment is needed to establish how the approach performs in routine practice and whether it improves clinical outcomes.

Abstract

Purpose: To evaluate Sybil, a MONAI-based nodule model, and their combination for one-year lung cancer detection in NLST and P-IELCAP. Methods: Baseline LDCT examinations were analyzed; the prespecified OR-stacked model combined Sybil, MONAI nodule detection, and a safety-net rule. Discrimination and operating-point performance were assessed. Results: Among 5,110 examinations, 65 lung cancers were diagnosed within one year. Sybil AUC was 0.93 in both cohorts; MONAI AUCs were 0.90 in NLST and 0.79 in P-IELCAP. The composite model detected 63/65 cancers (sensitivity 96.9%) and classified 2,731/4,965 non-cancer examinations as negative (specificity 55.0%). Conclusion: The composite AI model achieved high sensitivity while maintaining moderate specificity.

Status: Submitted to Radiology: Cardiothoracic Imaging. Manuscript RYCT-26-0087.R3. The full manuscript is not attached.

Download the abstract

Download Abstract (PDF)