EAGLE AI model detects oesophageal cancer on routine chest CT with 90.0% sensitivity

Researchers reported in Nature Medicine on 22 September 2026 that an AI model called EAGLE can detect oesophageal cancer and precancerous lesions from chest non-contrast CT. The authors validated it on 80,612 patients across 12 centres in three countries ¹.

What the study found

The authors trained the Esophageal AI-Guided malignant Lesion Evaluation (EAGLE) model on 6,813 patients from two centres ¹. They describe detecting these lesions on non-contrast CT as "a task historically considered impossible", because the oesophagus is a hollow structure prone to collapse and motion artefacts ¹. Validation covered both opportunistic screening on existing scans and population-based screening ¹.

The abstract reports results across several cohorts:

• External test cohorts for opportunistic screening (eight centres, 11,466 patients): 98.5% specificity, 90.0% sensitivity for cancer and 52.5% sensitivity for precancerous lesions ¹.
• Low-dose CT validation (two centres, 1,607 patients): performance comparable to the main external tests ¹.
• A real-world calibration cohort (three centres, 35,402 patients): calibration cut false positives by 72.7% while preserving sensitivity ¹.
• Prospective hospital validation (17,446 patients): a positive predictive value of 42.2% ¹.
• Real-world low-dose screening (10,959 patients): 99.94% specificity ¹.

The authors also compared CT with endoscopy in paired cohorts at two centres (702 patients). At a higher-sensitivity operating point, sensitivity was 65.0% for precancerous lesions and 78.4% for stage I cancer ¹. Exploratory analyses of a prospectively enrolled cohort suggest that referring high-risk individuals for endoscopy could improve screening efficiency ¹.

What it means for the market

The commercial case rests on reuse of scans already taken. The model reads chest CT acquired for other reasons, so a health system would add software to existing imaging rather than a new test for patients ¹. For buyers of radiology AI, including NHS trusts, this would be an additional finding from studies they already report.

The low-dose results matter most for screening programmes. The authors say the low-dose validation supports "EC screening through lung-cancer screening programs" ¹. A system that already runs low-dose lung screening could, on this evidence, look for a second cancer in the same images.

The positive predictive value sets the operational cost. In the prospective hospital cohort, 42.2% of positive flags were confirmed, so more than half of referrals would be false alarms ¹. Each positive flag implies a follow-up investigation, and the pathway the authors explore is referral to endoscopy ¹. Endoscopy demand is therefore the main downstream consequence for any buyer.

The abstract does not report several things buyers and investors would need ¹. It gives no cost or cost-effectiveness data. It does not name a developer, commercial product or regulatory status. It does not address UK pathways such as community diagnostic centres or the NHS lung screening programme. Procurement relevance for UK and European buyers therefore remains unknown on this evidence.

Caveats

The abstract does not name the three countries, and no UK site is identified ¹. Sensitivity for precancerous lesions was 52.5% in the main external tests ¹. The 78.4% stage I figure came from a higher-sensitivity setting, and the abstract does not give the specificity at that setting ¹.

Performance depended on calibration. False positives fell by 72.7% only after the model was calibrated on a 35,402-patient real-world cohort ¹. The finding on screening efficiency comes from exploratory analyses ¹. The abstract reports detection accuracy, not effects on stage at diagnosis or survival. The authors themselves say EAGLE "has the potential to serve as a scalable tool", not that it is one ¹.