Summary
A Nature Medicine study developed EAGLE, an AI model that identifies esophageal cancer and precancerous lesions on noncontrast chest CT scans. It was validated across 80,612 patients in 12 centres and three countries, including routine and low-dose CT settings.
A study published in Nature Medicine describes an artificial-intelligence model that identifies esophageal cancer and high-grade intraepithelial neoplasia (HGIN), a precancerous lesion, from chest CT scans taken without contrast dye. Called Esophageal AI-Guided malignant Lesion Evaluation, or EAGLE, the model was trained on 6,813 patients and validated across 12 centres in China, the Czech Republic and Australia, involving 80,612 patients in hospital and screening settings.
In external testing on existing regular-dose noncontrast CT scans, EAGLE achieved 98.5% specificity and 90.0% sensitivity for esophageal cancer. Its sensitivity for HGIN was 52.5%. The results indicate that routine chest scans could potentially provide an additional opportunity to identify esophageal disease without requiring a separate imaging examination.
How EAGLE analyses a chest CT scan
Noncontrast CT is widely used because it is rapid, noninvasive and does not require an injected contrast agent. However, early esophageal lesions can be very small, while the esophagus is a narrow tube that can collapse and move with the heart and nearby blood vessels. These features make subtle abnormalities difficult to distinguish from normal tissue, including for experienced radiologists.
EAGLE uses a two-stage system. First, it locates the esophagus within the three-dimensional CT scan. It then analyses that region to classify the patient as positive or negative, outline suspicious lesions with segmentation masks and generate heatmaps showing the areas that influenced its decision. The training data included pathology-confirmed cancers and HGIN, along with expert annotations of lesion locations.
The operating threshold can be adjusted for different clinical uses. For opportunistic screening among general hospital patients, the researchers selected a threshold targeting 99% specificity. For use before endoscopy in high-risk screening populations, they selected a higher-sensitivity threshold targeting 98% sensitivity.
Performance in hospital and screening settings
The regular-dose external test included 11,466 patients from eight centres. EAGLE’s overall area under the receiver operating characteristic curve was 0.976, with 89.5% sensitivity and 98.5% specificity. Sensitivity for stage I esophageal cancer was 60.1%, compared with 90.0% for esophageal cancer overall and 52.5% for HGIN.
In a reader study involving 17 radiologists and 300 CT scans, assistance from EAGLE increased sensitivity for esophageal cancer from 71.9% to 85.7% and specificity from 79.6% to 91.7%. The improvement was also seen for early lesions, with the average reader sensitivity increasing by 16.6 percentage points for HGIN and 22.7 percentage points for stage I cancer.
The researchers also adapted the system for low-dose CT, including scans used in lung-cancer screening. In a validation cohort of 1,607 scans, an LDCT-trained version achieved 88.4% sensitivity at an operating point corresponding to 99.0% specificity. In a separate real-world cohort of 10,959 people aged 45 to 75 years, EAGLE-Plus produced eight positive predictions and achieved 99.94% specificity. One of those cases had not been reported as suspicious initially and was confirmed as esophageal cancer eight days later.
A prospective hospital validation processed 17,446 patients between January and April 2025, with follow-up completed on 31 July 2026. Forty-one esophageal cancers and six HGIN lesions were confirmed. The model produced 90 positive predictions, including 38 confirmed malignant cases, giving a positive predictive value of 42.2%. Sensitivity was 87.8% for esophageal cancer and 80.9% for esophageal malignancy including HGIN.
The study also examined whether EAGLE could help prioritise people for endoscopy. In a paired CT–endoscopy cohort of 702 patients from two Chinese screening centres, the model detected HGIN with 65.0% sensitivity and stage I cancer with 78.4% sensitivity at the higher-sensitivity operating point. A separate exploratory analysis used data from 530 prospectively enrolled screening participants and added three resampled cancer cases to create a hybrid cohort of 533 people. In that simulation, referring only EAGLE-positive participants for endoscopy increased the estimated malignant-lesion detection rate from 1.7% to 5.2% and could have avoided 67.6% of endoscopic examinations. Two HGIN cases were missed in the simulated negative group.
The evidence is strongest for showing technical performance and workflow potential. The pre-endoscopy triage result is based on retrospective paired CT–endoscopy data and a simulation from one prospectively enrolled screening site, while the hospital and low-dose real-world validations had follow-up periods shorter than two years and incomplete adherence to recommended endoscopy. The authors also identify a need for broader validation in populations with more distal esophageal cancer and adenocarcinoma, and in female patients. The reported study endpoints concern detection, specificity and screening efficiency; effects on treatment outcomes or survival were not measured.