EAGLE AI finds early oesophageal cancer on routine chest CT — a task doctors long called impossible

Early oesophageal cancer is hard to catch. Endoscopy works, but it is invasive and hard to scale. Ordinary chest CT is everywhere — yet doctors long treated the food pipe on those scans as almost unreadable for tiny early lesions.

On 22 September 2026, Nature Medicine published EAGLE — Esophageal AI-Guided malignant Lesion Evaluation. The open-access paper says the model detects oesophageal cancer and high-grade intraepithelial neoplasia, a serious precancerous lesion, from noncontrast chest CT. The authors call that a task historically considered impossible. The gullet collapses. Early spots are tiny. Motion from the heart muddies the picture.

The team trained on 6,813 patients from two Chinese centres. They then tested across 12 centres in China, the Czech Republic, and Australia — 80,612 patients in all. Lead clinical sites include Sun Yat-sen University Cancer Center and Sichuan Cancer Hospital. Co-authors also include researchers at Alibaba’s DAMO Academy and Hupan Lab. Several Alibaba employees own company stock as part of standard pay; the paper discloses that. A sample portal is at eagle.damomed.com; it is not a substitute for the journal article.

On existing CT scans used for opportunistic screening, eight external centres (11,466 patients) saw 98.5% specificity, 90.0% sensitivity for cancer, and 52.5% sensitivity for precancerous lesions. Low-dose CT checks looked similar after a simulation training step — useful because lung-cancer screening already images the full gullet. In a reader study with 17 radiologists, EAGLE alone beat every reader on detecting those lesions. With EAGLE’s help, average reader sensitivity rose and false alarms fell. A calibrated version, EAGLE-Plus, ran prospectively on 17,446 hospital patients and posted a 42.2% positive predictive value among its 90 flags. In a real-world low-dose CT physical-exam cohort of 10,959 people, specificity reached 99.94%. Trial registry: ChiCTR2300074806.

Hope needs a brake. The paper’s own limits matter. More data are still needed on adenocarcinoma, female patients, and sites outside China. Some prospective arms had short follow-up, and not everyone offered endoscopy took it. Code is patent-protected; this is not an open-weights release. Opportunistic “finds” on scans done for other reasons are promising triage — not a finished population screening programme on their own. Keep this distinct from Stanford’s virtual-biotech drug-design story: different lab, organ, and modality.

The pattern is amplifier. Point a two-stage CT model at a modality people wrote off for early gullet disease, and it surfaces lesions humans miss — because that is what the system was built to look for.