CT scans may carry a second clue.

The human problem

Esophageal cancer is often difficult to see on ordinary chest CT and can be diagnosed after disease has advanced.

The prior constraint

Chest CT scans taken for other reasons contain the esophagus, but subtle lesions can be overlooked and endoscopy is not performed for everyone.

AI’s actual role

A deep-learning system analyzed non-contrast chest CT scans for signs of esophageal cancer and high-grade intraepithelial neoplasia, including scans acquired for other reasons.

The documented result

Across 11,466 patients at eight external centres, EAGLE flagged 90.0% of known cancers (sensitivity) and correctly left 98.5% of patients without cancer unflagged (specificity). It flagged 60.1% of stage I cancers and 52.5% of precancerous lesions. In a separate prospective cohort of 17,446 patients at one hospital, it flagged 87.8% of known cancers. Of 90 alerts, 38 were confirmed malignant—36 cancers and two high-grade intraepithelial neoplasias—so 42.2% of alerts were confirmed malignant (positive predictive value).

Why it may matter

If broader validation supports it, opportunistic review could draw attention to a hard-to-see disease on CT scans people already receive, without adding another scan.

Limitations

The prospective cohort had incomplete follow-up: among 52 people without confirmed malignancy, 28 had other findings, 19 were negative or had no symptoms, and five were lost to follow-up.

The 10,959-person low-dose CT cohort was retrospective and single-site. Eight people were flagged, one cancer was confirmed, and five did not undergo endoscopy, so its 12.5% positive predictive value is not a complete screening outcome.

External sensitivity for stage I cancer was 60.1% and for precancerous lesions was 52.5%; the study does not show reduced mortality or better patient outcomes.

The authors call for broader international validation. DAMO Academy (Hupan Lab) was the main funder, and eight authors reported Alibaba employment and stock compensation.

Unresolved questions

Will performance hold across more countries, cancer subtypes, scanners and care settings?

Can a complete follow-up pathway show whether alerts lead to earlier diagnosis and better outcomes without excessive procedures?

How should clinicians manage uncertain positives and precancerous lesions?