Performance of Sybil AI for lung cancer risk assessment: a head-to-head comparison of cardiac vs. lung CT scans
M MozafaryBazargany, A Hashemi, A Reeves, K Atlas, C Zhang, R Mirjalili, A Azimi, A Ghaffari Jolfayi, H Ghoshouni, H Mohammadi, H Zarei, Z Fayad, D Yankelevitz, N Wong, M NaghaviAbstract
Background
Sybil is a validated deep-learning model designed to predict 1-to-6-year lung cancer (LC) risk from a single low-dose computed tomography (LDCT) scan without clinical data or radiologist annotations. While Sybil was developed for screening-intent low-dose lung CTs, millions of diagnostic (regular dose) lung scans and coronary artery calcium (CAC) scans are performed annually on patients who are not eligible for lung cancer screening. Studies have shown that about two-thirds of patients diagnosed with lung cancer are not eligible for lung cancer screening using existing criteria. The cardiac scans typically omit the upper third of the lungs, where approximately 50% of lung cancers occur. This study evaluates whether Sybil maintains predictive accuracy on CAC scans despite this restricted field of view.
Methods
We analyzed 7,204 participants from the Framingham Heart Study (FHS, n=2,415) and the Multi-Ethnic Study of Atherosclerosis (MESA, n=4,789) who underwent both LDCT and CAC scans. Sybil’s predictive power was measured using the Area Under the Receiver Operating Characteristic Curve (AUC) for 1-year to 6-year LC risk. We compared performance between full-lung LDCT and truncated CAC scans to determine if the AI identifies systemic patterns in visible lung tissue that compensate for missing apical segments.
Results
Sybil demonstrated comparable predictive power across modalities. In the FHS cohort, the AUC for 1-year LC prediction was 0.88 (95% CI: 0.81-0.95) on CAC scans versus 0.92 (95% CI: 0.87-0.97) on full lung scans. In the MESA cohort, the 1-year AUC remained robust at 0.81 for CAC scans. Despite missing approximately 33% of the superior lung volume, the AI successfully identified architectural features and predictive patterns throughout the mid-to-lower lung parenchyma as reliable indicators for future malignancy. No statistically significant degradation in predictive performance was observed between truncated cardiac scans and full-field lung scans (p > 0.05), suggesting the model captures systemic pulmonary risk signals rather than relying solely on localized lesion detection.
Conclusion
Sybil AI achieves comparable lung cancer risk prediction on cardiac CAC scans and full lung CTs by detecting systemic patterns in visible lung regions. This enables opportunistic screening using millions of annual CAC scans to identify high-risk individuals currently missed by traditional lung cancer screening criteria.