Esophageal cancer has long been one of the hardest malignancies to catch early. Symptoms tend to appear late, and the screening tools that exist are not easily scaled to whole populations. A study published in Nature Medicine on 22 September 2026 describes a possible change in that calculus: an artificial intelligence model that detects esophageal cancer and precancerous lesions on noncontrast chest CT, a type of scan already performed in enormous numbers around the world.
The model is called Esophageal AI-Guided malignant Lesion Evaluation, or EAGLE. Its developers describe the underlying task — finding early esophageal malignancies on noncontrast computed tomography — as one that had historically been considered impossible. Their study, published under DOI 10.1038/s41591-026-04656-4, reports training and validation across thousands of patients and more than a dozen clinical centers in three countries.
The screening gap EAGLE tries to close
According to the study's abstract, the absence of accurate, noninvasive and scalable screening tools has kept early esophageal cancer detection a persistent global health challenge. Existing approaches have not delivered a practical way to screen large populations, leaving many cases to be discovered only after symptoms emerge.
Noncontrast CT would seem like an obvious candidate for a solution, because the scans are widely accessible and frequently obtained for lung and other chest indications. But the esophagus is a difficult target. As a hollow tubular structure, it is prone to collapse and to motion artifacts, and small early malignant lesions can be hard to distinguish from normal tissue. That anatomy and imaging physics are why the authors characterize the detection task as one that has been considered out of reach rather than merely difficult.
EAGLE was built to address exactly that problem: identifying both cancer and precancerous lesions from chest noncontrast CT images.
Training and validation at scale
The study's development pipeline is notable for its size and geographic spread. EAGLE was trained on 6,813 patients drawn from two centers. It was then validated across 12 centers in three countries, involving 80,612 patients in both opportunistic and population-based screening settings.
Those two settings matter. Opportunistic screening means applying the model to CT scans that were already taken for other clinical reasons, without requiring anyone to undergo a new procedure. Population-based screening refers to organized programs that invite a defined population for testing. The validation therefore tested the model in circumstances that resemble how a real screening tool would be used, rather than only in a curated research dataset.
- Training cohort: 6,813 patients from two centers.
- Validation: 12 centers across three countries, involving 80,612 patients.
- Opportunistic-screening external test cohorts: eight centers, n = 11,466.
- Low-dose CT validation: two centers, n = 1,607.
- Real-world calibration cohort: three centers, n = 35,402.
How the model performed
In the opportunistic screening setting, the multicenter external test cohorts — eight centers with 11,466 patients — produced 98.5% specificity. For cancer, sensitivity reached 90.0%. For precancerous lesions, sensitivity was 52.5%.
Validation on low-dose CT, covering two centers and 1,607 patients, showed performance comparable to the standard noncontrast CT results. That finding is significant because it suggests esophageal cancer screening could potentially be layered onto lung cancer screening programs, which already rely on low-dose CT and already reach large numbers of people.
The researchers also report calibration in a real-world cohort spanning three centers and 35,402 patients, a step that matters because a model's discrimination scores do not guarantee that its predicted probabilities align with observed outcomes in routine practice.
Reading the numbers carefully
The headline figures reward close reading. Specificity of 98.5% means that in the external test cohorts, very few people without disease would be flagged — a crucial property in screening, where false positives generate anxiety, follow-up procedures and cost. A 90.0% sensitivity for cancer indicates the model caught the large majority of cancers in that setting.
The 52.5% sensitivity for precancerous lesions tells a more complicated story. Precancerous changes are, by definition, earlier and subtler than invasive cancer, and in the external test cohorts the model missed close to half of them. That does not erase the tool's potential value — identifying a substantial share of early, treatable lesions could still shift outcomes — but it does mean EAGLE, as described, is not a perfect early-warning system.
Why lung cancer screening programs matter here
The low-dose CT results may be the most practically consequential part of the study. Lung cancer screening already brings high-risk individuals into CT scanners on a recurring basis. If a model can extract esophageal cancer signals from the same images without additional radiation, contrast agents or appointments, the marginal cost of adding that check could be far lower than building a separate screening pathway from scratch.
The authors frame the low-dose CT validation as support for screening esophageal cancer through lung cancer screening programs. That is a logistical argument as much as a technical one: the infrastructure, the patient population and the imaging hardware are already in place.
What the study does not settle
Multi-center validation across three countries is a meaningful step toward generalizability, but important questions remain outside what a single validation study can answer. How the model performs prospectively in a live screening program, how its outputs integrate with radiologist workflows, and whether the reported specificity holds across populations with different risk profiles are all open issues.
The gap between cancer sensitivity and precancerous-lesion sensitivity also points to a design tradeoff. A tool that reliably flags invasive disease could still be useful for triage, even if it functions better as a complement to existing diagnostic pathways than as a standalone filter.
The larger significance of the work is conceptual. If noncontrast chest CT — an imaging study already embedded in routine care — can carry usable information about esophageal malignancy, then the question shifts from whether screening is technically possible to how such a model should be deployed, monitored and validated over time. EAGLE offers an early, large-scale answer to the first question and a detailed agenda for the next ones.
This article is based on reporting by Nature Medicine. Read the original article.
Originally published on nature.com
