AI Model EAGLE Detects Esophageal Cancer on Routine Non-Contrast CT, with Accuracy Validated in 11,000 Patients from China, Czechia, and Australia

AI Model EAGLE Detects Esophageal Cancer on Routine Non-Contrast CT, with Accuracy Validated in 11,000 Patients from China, Czechia, and Australia

On September 22, Nature Medicine published a study on EAGLE, a model that detects cancer and precancerous changes in the esophagus on non-contrast CT. Validated across eight centers in three countries (11,466 patients), the model identified 89.5% of cancer cases while raising a false alarm in only 1.5% of healthy individuals. The system is already in use at three Chinese hospitals, where it has confirmed cancer cases missed by conventional diagnosis.

Esophageal cancer is almost always found late: approximately 511,000 new cases and 445,000 deaths occur worldwide each year, with five-year survival at 36.9% in China and 18.5% in the United States. No population-wide screening exists. Endoscopy, which involves passing a camera through the throat, is invasive and is offered only in select regions of China for high-risk groups.

Millions of people already undergo non-contrast chest CT for various reasons, and these scans capture the entire esophagus; such scans account for 40% of all CT examinations worldwide. The esophagus is a narrow tube subject to cardiac and vascular motion, making early changes difficult for radiologists to discern. Among most patients who died of esophageal cancer, CT reports from the preceding five years contained no suspicious findings: the signal had been sitting in the archive for years, but no one was reading it.

A team from Alibaba DAMO Academy (a division of Alibaba) and Chinese oncology centers built EAGLE in two stages. Because the tumor occupies a tiny fraction of the total scan volume, one neural network first locates and segments the esophagus, then a second network searches within it for malignant or precancerous changes (abnormal cells that have not yet become cancer but have the potential to progress) and marks them on the image. The model was trained on voxel-level tumor annotations (voxels are three-dimensional equivalents of pixels): boundaries that radiologists draw on contrast-enhanced CT were transferred to the non-contrast scan through image registration. This level of precision allowed the model to distinguish barely visible early signals.

EAGLE outperformed each of the 17 radiologists individually, and with its assistance, radiologists detected cancer in 85.7% of patients compared with 71.9% without it. For stage I cancer, the model detects notably fewer cases (60 to 66%) compared with 89 to 90% for more established tumors.

The model was also fine-tuned on lower-dose scans of the type used in lung cancer screening programs for smokers; smoking is a shared risk factor for both cancers. The model still identified 88.4% of cancer cases, meaning a single scan intended for the lungs can also detect esophageal cancer without an additional procedure.

The model is already in clinical use. In the first wave across three Chinese hospitals (21,000 patients), it found three cancer cases missed by conventional diagnosis, one of them 21 months before the official diagnosis was made. After retraining on these cases, the model reduced false positives in the second wave at the same hospitals (14,600 patients) by 72.7%. In an endoscopic screening program where EAGLE triages high-risk patients, the proportion of confirmed findings at endoscopy rose from 1.7% to 5.2%, and the number of endoscopies per confirmed finding dropped from 59 to 19.

The same group has been developing this approach since 2022, when it first demonstrated on 180 patients at a single center that detecting esophageal tumors on non-contrast CT was feasible. Over four years, this has grown into validation across 11,000 patients in three countries and clinical practice covering tens of thousands in China.

The authors suggest that the same annotation technique could open a path to detecting other cancers on non-contrast CT that have long been beyond its reach. If so, the archive of existing scans could become a source of new diagnoses without invasive procedures.