AI model reads routine chest CT scans for oesophageal cancer, Nature Medicine study reports

A team led by researchers in China, working with Alibaba's DAMO Academy and including a co-author in Prague, has reported an AI model that looks for oesophageal cancer and precancerous lesions on ordinary chest CT scans taken without contrast dye. The paper was published in Nature Medicine on Tuesday 22 September. The authors describe detection from these scans as "a task historically considered impossible".

The problem they set out is that oesophageal cancer lacks "accurate, noninvasive, scalable screening tools", and that on non-contrast CT the oesophagus is "a hollow tubular structure prone to collapse and motion artifacts". The paper puts the global toll at an estimated 511,000 new cases and 445,000 deaths in 2022.

What the study did

The model, called EAGLE (Esophageal AI-Guided malignant Lesion Evaluation), was trained on 6,813 patients from two centres. It was then tested in several settings, including existing hospital scans, low-dose CT from lung-cancer screening, a real-world deployment and a prospective hospital study. The external test centres were in China, the Czech Republic and Australia.

The headline results, as stated in the paper:

• On existing scans at eight external centres (11,466 patients), EAGLE "achieved 98.5% specificity, with 90.0% sensitivity for cancer and 52.5% for precancerous lesions".
• On low-dose CT at two centres (1,607 patients) it "showed comparable performance", which the authors say supports screening through lung-cancer programmes.
• Calibration on a real-world cohort (35,402 patients across three centres) "reduced false positives by 72.7% while preserving sensitivity".
• A prospective hospital validation of 17,446 patients "achieved a 42.2% PPV" (positive predictive value).
• In real-world low-dose screening of 10,959 people, specificity reached 99.94%.
• In a 17-reader comparison on 300 scans, the paper says EAGLE "outperformed all 17 readers" in identifying precancerous lesions and cancer.

The prospective numbers in detail

In the prospective study, run at one Shanghai centre from 1 January to 30 April 2025 using a recalibrated version called EAGLE-Plus, the model flagged 90 of 17,446 patients, and 38 of those were confirmed as cancer or high-grade precancerous lesions. Of the 90, 48 had already been managed by the standard pathway, and the multidisciplinary team referred 16 of the rest for endoscopy. Three of those 16 completed endoscopy, which found one cancer and two negative cases.

Separately, in a retrospective real-world low-dose cohort of 10,959 people screened from July to December 2024, the paper reports a positive predictive value of 12.5%, alongside 99.94% specificity.

On pre-endoscopy use, the abstract says of a separate, smaller screening cohort: "Exploratory analyses of a prospectively enrolled cohort suggest that referring high-risk individuals for endoscopy could improve screening efficiency."

What the authors say is not yet known

The paper lists several limitations. It says broader international validation "remains essential", particularly where adenocarcinoma and cancers lower in the oesophagus are more common. It says observed sex-based differences in sensitivity may reflect an imbalance in the training data, and calls for more annotated data from female patients. It says the value of EAGLE before endoscopy "requires larger and more mature prospective data"; the paper notes the present evidence rests on retrospective cohorts and a simulation from a single site with limited positive cases. Follow-up in the prospective hospital study and the real-world low-dose cohort was under two years, and compliance with recommended endoscopy was suboptimal, so the authors say "the model's true sensitivity may not be fully reflected". They also say recall bias in the reader study "cannot be completely excluded".

Eight of the authors are employees of Alibaba Group and own Alibaba stock, according to the paper's competing-interests statement. The paper says the work was supported by DAMO Academy (Hupan Laboratory), with some authors also holding provincial research grants. The authors conclude that EAGLE "has the potential to serve as a scalable tool for early EC screening".

On our reading, the retrospective accuracy figures are strong, but the prospective referral numbers are small, and the authors themselves say its value as a pre-endoscopy screening test still needs larger and more mature prospective data.