- DAMO Academy's EAGLE model detects esophageal cancer from non-contrast chest CT with 90% sensitivity and 98.5% specificity in external validation across eight centers.
- The model was trained on 6,813 patients at two Chinese cancer centers and validated on more than 80,000 patients across 12 centers in China, the Czech Republic and Australia.
- EAGLE detected 52.5% of precancerous high-grade intraepithelial neoplasia lesions, which are typically invisible on plain CT scans.
- Researchers state the AI is designed to identify high-risk individuals for follow-up endoscopy, not to replace endoscopic screening or diagnosis.
- Performance has not been established outside the studied populations, and whether earlier AI-flagged detection improves survival outcomes remains unreported.
Model Development and Validation
Alibaba's DAMO Academy, together with Sichuan Provincial Cancer Hospital and Sun Yat-sen University Cancer Center, developed the EAGLE (Esophageal AI-Guided malignant Lesion Evaluation) model to address the difficulty of detecting early-stage esophageal cancer on routine chest imaging. The esophagus is a narrow, collapsible organ whose early lesions are obscured by cardiac and vascular motion, making them nearly invisible to radiologists on non-contrast CT.
The team trained the model on CT scans from 6,813 patients at two Chinese cancer centers, using expert-annotated lesion masks paired with endoscopy reports and contrast-enhanced CT to teach the AI to recognize subtle abnormalities. The study was published in Nature Medicine on September 22, 2026.
Clinical Performance Across Cohorts
In an external validation set of 11,466 patients across eight centers, EAGLE detected 90.0% of esophageal cancers and 52.5% of precancerous high-grade intraepithelial neoplasia, with 98.5% specificity. Total validation covered 80,612 patients across 12 centers in three countries.
The model maintained comparable accuracy when applied to low-dose CT scans commonly used for pulmonary screening, evaluated in a cohort of 1,607 individuals. Following recalibration using real-world data from 35,402 patients, false positive rates declined by 72.7% while preserving sensitivity. During prospective evaluation involving 17,446 hospital patients, 42.2% of positive AI flags were confirmed as true malignancies.
Intended Role and Limitations
The researchers and collaborating clinicians emphasize that EAGLE is not a replacement for endoscopy. Wang Qifeng, chief physician at Sichuan Provincial Cancer Hospital, stated that the AI's role is to screen out high-risk individuals who should then undergo endoscopic examination, improving the precision and yield of invasive testing rather than eliminating it.
All training and validation cohorts were drawn from hospitals in China, with additional testing in the Czech Republic and Australia; performance in other populations has not been established. The study did not report whether earlier detection via AI flagging translates into improved survival outcomes. More than half of patients flagged by the model in prospective testing did not have confirmed cancer, meaning follow-up imaging or endoscopy remains necessary for all positive results.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01
