An artificial intelligence model detected 90% of esophageal cancers on plain chest CT scans, the ordinary kind done every day in hospitals, according to a study published Tuesday in Nature Medicine. The model, called EAGLE, was trained on 6,813 patients and tested across 12 centers in China, the Czech Republic and Australia covering 80,612 people, on a task the authors call “historically considered impossible.”
If you already qualify for a low-dose chest CT to screen for lung cancer, or you have had heartburn for years and a Barrett’s esophagus diagnosis, this is the finding to file away. That scan already includes your esophagus, and nobody currently reads it for early esophageal cancer. But EAGLE is research software running mostly in Chinese hospitals. You cannot request it in an American radiology department, and it does not replace an endoscopy your doctor thinks you need.
Key takeaways
- EAGLE caught 90.0% of esophageal cancers at 98.5% specificity across eight outside centers, but only 52.5% of precancerous lesions.
- In a live hospital rollout of 17,446 people, fewer than half the patients it flagged had cancer or a precancerous lesion.
- The study reports no data on whether earlier detection saves lives.
What the study found
Esophageal cancer is usually caught late. There were an estimated 511,000 new cases and 445,000 deaths from it worldwide in 2022, and five-year overall survival is 18.5% in the United States and 36.9% in China, meaning fewer than 1 in 5 American patients are alive five years after diagnosis. Endoscopy catches it early but is invasive, and few people volunteer.
In external testing at eight centers covering 11,466 patients, EAGLE reached 98.5% specificity and 90.0% sensitivity for cancer, correctly clearing 985 of every 1,000 people without cancer and catching 9 of every 10 who had it. Sensitivity for high-grade precancerous lesions was 52.5%, roughly a coin flip, and 60.1% for stage I cancer.
It also worked on low-dose CT, the scan used for lung cancer screening: in two centers covering 1,607 patients, a version tuned for low-dose images caught 88.4% of cancers. The same group published a sibling model last year, GRAPE, that reads the same kind of scan for stomach cancer.
In a reader study, 17 radiologists reading the same 300 scans caught 71.9% of cancers on their own. With EAGLE’s output in front of them, that rose to 85.7%, and specificity went from 79.6% to 91.7%. Both changes had less than a 1 in 1,000 chance of being coincidence.
Then they ran it live. In a prospective deployment at one Shanghai hospital covering 17,446 patients, EAGLE flagged 90 people. Thirty-eight had confirmed cancer or a precancerous lesion, a positive predictive value of 42.2%.
Dr. Kumar’s take
Pulling a small early tumor out of a collapsed, moving tube on a scan with no contrast dye is hard, and the reader study shows the model is seeing something trained radiologists were not. Two things still keep me from calling this a screening breakthrough for Americans.
The first is tumor biology. This work is overwhelmingly Chinese, where esophageal cancer is mostly squamous cell carcinoma in the middle of the esophagus. The authors note that the two main subtypes of this cancer differ in their risk factors, where they sit along the esophagus, and which parts of the world they turn up in. The paper hints at the problem: EAGLE did worse on lesions at the junction of the esophagus and stomach, and the authors list validation “in populations with a higher prevalence of distal EC and adenocarcinoma” as an open limitation.
The second is the gap between finding a cancer and helping a patient. A 42.2% positive predictive value means most people this model flags go on to worry, and some to a sedated endoscopy, for something that is not cancer. Sensitivity for the precancerous lesions you would remove to prevent a death was 52.5%. And the paper reports no data on whether running EAGLE on people’s scans changes how many of them die of this cancer. That is the trap multi-cancer blood tests are sitting in: impressive detection numbers, no demonstration yet that detection means fewer deaths. The authors do not oversell it, writing only that EAGLE has “the potential to serve as a scalable tool for early EC screening.”
What it means for you
Nothing to do today. No American hospital is running this on your scan.
What is worth doing already has evidence behind it. If you are 50 to 80 with a 20 pack-year smoking history, you qualify for annual low-dose CT lung cancer screening, and smoking raises esophageal cancer risk too. That scan is the decision that matters, and it puts an image of your esophagus in your chart either way.
If you have had reflux for years, or you have Barrett’s esophagus, that is a conversation with a gastroenterologist about surveillance, not a reason to ask for a CT. Endoscopy remains the test that finds and removes precancerous tissue.
If AI screening reaches your hospital, the question to ask is not how many cancers it finds. It is what happens to the people it flags who turn out to be fine.