EAGLE AI Shows High Specificity for Esophageal Cancer Screening

An artificial intelligence (AI) model designed to detect esophageal cancer (EC) and precancerous lesions on routine noncontrast CT (NC CT) scans showed high specificity in multicenter opportunistic screening cohorts, according to a study published in Nature Medicine.

The investigators noted that detecting EC on NC CT has been challenging because the esophagus is a hollow tubular structure prone to collapse and motion artifacts, making small early lesions difficult to distinguish from normal tissue. The Esophageal AI-Guided malignant Lesion Evaluation (EAGLE) model was trained on scans from 6,813 patients at two centers and validated across 12 centers in China, the Czech Republic, and Australia, involving 80,612 patients in opportunistic and population-based screening settings.

In external test cohorts from eight centers (n = 11,466), EAGLE achieved 98.5% specificity, with 90.0% sensitivity for EC and 52.5% for high-grade intraepithelial neoplasia (HGIN). Sensitivity for stage I EC in these cohorts was 60.1%. In a reader study, EAGLE assistance improved the sensitivity of 17 radiologists from 71.9% to 85.7% and specificity from 79.6% to 91.7% (both P < 0.001).

A calibrated version, EAGLE-Plus, reduced false positives by 72.7% in a real-world cohort while preserving sensitivity. In prospective hospital validation (n = 17,446), it achieved a positive predictive value (PPV) of 42.2%; in retrospective real-world low-dose CT (LDCT) validation (n = 10,959), it achieved 99.94% specificity with a PPV of 12.5%. In a retrospective paired CT–endoscopy cohort (n = 702), the original EAGLE model, set to a higher-sensitivity threshold, had sensitivities of 65.0% for HGIN and 78.4% for stage I EC.

The authors reported that exploratory analyses suggest referring EAGLE-identified high-risk individuals for endoscopy could improve screening efficiency, but they noted that this evidence rests on retrospective cohorts and a single-site simulation. Other limitations included follow-up of less than 2 years in the prospective hospital and real-world LDCT validations, suboptimal adherence to recommended endoscopy, and the need for broader international validation.

Source: Zhou J, Guo G, Yao J, et al. Large-scale esophageal cancer screening through noncontrast computed tomography and artificial intelligence. Nat Med. Published online September 22, 2026. doi:10.1038/s41591-026-04656-4