Non-invasive esophageal cancer screening: DAMO Academy starts with a lung CT scan.

Wang Qifeng frequently visits district and county hospitals to provide guidance.

This chief physician of the radiotherapy department at Sichuan Provincial Cancer Hospital has seen many regrettable cases at the grassroots level. Patients often arrived unable to eat, missing the opportunity for radical treatment. Checking the hospital's records, it was discovered that this person had been hospitalized two years prior for pneumonia and had undergone a chest CT scan.

Comparing the images from two years ago with those from now, the esophageal wall had already thickened back then.

At the time, no one noticed that the patient did not mention any problems with eating, and the radiologist only looked at the condition of the lungs.

On September 24, Alibaba DAMO Academy, in collaboration with Sichuan Cancer Hospital, Sun Yat-sen University Cancer Center, and other institutions, released DAMO EAGLE, an AI model for esophageal cancer screening. The related paper was published in *Nature Medicine*. In the initial screening stage, the model directly reads plain CT scans without intubation or contrast agent injection. Even after the AI indicates a positive result, patients still need to undergo endoscopy and pathological examination for a definitive diagnosis.

A paper related to DAMO EAGLE was published in *Nature Medicine*.

Its purpose is to help doctors examine any esophageal effects captured during lung scans.

This is the fourth cancer screening model released by DAMO Academy, following those for pancreatic cancer, gastric cancer, and colorectal cancer. While the gastric cancer model GRAPE, released last year, is used before gastroscopy, EAGLE starts by looking for esophageal cancer risk from chest CT scans already taken at the hospital.

01 The gold standard is there, but there are too few people who can perform endoscopy.

China faces a significant burden of esophageal cancer globally. Data from the National Cancer Center shows that in 2024, China had approximately 200,000 new cases of esophageal cancer and about 172,000 deaths, accounting for nearly half of the global total.

In 2024, the number of deaths from major cancers in China (in ten thousand people) was 172,100 for esophageal cancer.

The high incidence rates are concentrated in the Taihang Mountains region of Henan, Hebei, and Shanxi, as well as the Huai River basin, Sichuan, Guangdong, and Fujian, with rural areas showing a significantly higher rate than urban areas. Wang Qifeng said this is related to the locals' preference for eating very hot food and drinking very hot tea.

Approximately 70% of patients are diagnosed at an advanced stage.

The prognosis differs greatly between early and late stages. When the lesion is still in the mucosal layer, a single endoscopic minimally invasive surgery can cure it, with a 5-year survival rate of over 95%. However, once metastasis occurs, Wang Qifeng estimates that regardless of the treatment method used, the 5-year survival rate is only about 10% to 15%.

It usually takes only about six months for the condition to progress from a feeling of something stuck in the throat when swallowing to being able to only drink liquids.

Wang Qifeng has a vivid memory of this process. When he was a sophomore in medical school, his grandmother was diagnosed with esophageal cancer in Changzhi, Shanxi Province, which is also a high-incidence area. At the time of diagnosis, the lesion was already about 5 centimeters in size. After the first stage of treatment, his grandmother was able to eat again, but eventually developed esophageal leakage and passed away only 8 months after diagnosis.

There are ways to detect it early; upper gastrointestinal endoscopy is the recognized gold standard. The problem is that we can't do it all.

According to official estimates, there are more than 100 million people in China who are at high risk of esophageal cancer and need screening. Approximately 30 million upper gastrointestinal endoscopy examinations are actually performed each year.

The shortage comes from several areas. It takes 3 to 5 years to train an endoscopist, and each examination takes 20 to 30 minutes. A regular endoscopy costs two to three hundred yuan, while the painless endoscopy that people prefer costs around 1000 yuan. Then there's the fear—a tube about 1 centimeter thick going from the mouth to the stomach. Wang Qifeng says even his own parents and relatives are afraid.

Areas with high incidence rates are often also the places with the fewest endoscopists.

02 The most difficult organ to identify, because even doctors can't label it.

Plain CT scans are one of the most commonly used imaging examinations in China, performed in outpatient clinics, emergency rooms, physical examinations, and hospitalizations, generating hundreds of millions of images annually. The scanning range of a chest CT scan inherently covers the esophagus.

However, it was generally believed in the industry that it was almost impossible to find early-stage esophageal cancer using plain CT scans.

The esophagus is a muscular tube about 25 centimeters long. It is normally closed, but it moves with peristalsis when speaking or swallowing. It is also close to the heart and major blood vessels, so breathing and heartbeat can change its shape. Early lesions are limited to the epithelial or mucosal layer, and the changes in thickness and density on CT scans are very small.

Guo Guangyu, an algorithm expert at DAMO Academy, used an analogy: it's like looking for a small ball of fluff on a wrinkled pair of trousers.

Wang Qifeng compared the pancreas to this, noting that the pancreas is a solid organ, and even a slight abnormality can appear on imaging. The esophagus, on the other hand, only shows a fleeting glimpse of its shape on a CT scan, and changes exceeding 5 millimeters are considered abnormal.

Model training is difficult even in the first step of annotation. Doctors cannot see early lesions on plain CT scans, so there is no way to annotate the training data.

The team's approach was to use other examinations to locate the lesion. First, endoscopic surgery and pathology reports were used to determine the centimeter-level location of the lesion. Then, an esophageal radiologist meticulously marked the lesion on an enhanced CT scan. Finally, through image registration, the markings were transferred to the same patient's plain CT scan. Each case was marked by two doctors, and in cases of disagreement, a doctor with eight years of experience reviewed the markings.

The second problem lies with low-dose CT scans. Low-dose CT scans are often used for physical examinations of lung nodules, but it's difficult to accumulate enough low-dose CT scans for esophageal cancer patients. Guo Guangyu explained that once a patient discovers an abnormality, they will undergo a standard-dose CT scan at the hospital.

The team then set up a "virtual CT scanner," put in a CT scan with a normal dose to reduce the dose, and simulated low-dose images to train the model.

The model works in two steps. First, it segments the esophageal region from the entire CT scan, eliminating interference from surrounding structures. Then, it performs a detailed analysis of the esophagus, providing patient-level benign/malignant assessments, lesion outlines, and thermal maps. Doctors can see where the AI is looking.

DAMO EAGLE identified a T1 stage esophageal cancer lesion on a plain CT scan. From left to right: plain CT scan, doctor's annotation, model segmentation results, and heat map.

In early 2024, the model underwent its first large-scale real-world retrospective study, with a sensitivity of around 85% and a specificity of approximately 98%.

The current version has been validated in 3 countries, 12 centers, and over 80,000 patients. The paper shows that in opportunistic screening scenarios, the model has a sensitivity of nearly 90% for esophageal cancer and 52.5% for precancerous lesions with weaker signals. After real-world iterations, the specificity in clinical scenarios is 99.2%.

The low-dose version had a sensitivity of 88.4%, which was essentially equivalent to the standard-dose version. In 10,959 asymptomatic individuals undergoing routine checkups, the specificity was 99.94%.

Guo Guangyu has a very direct conversion for the specificity figure: "A false positive means that a patient may have to spend more than 1,000 yuan to do another test."

The tools included in the physical examination process are mostly used by healthy people, and the false alarm rate determines whether they can be truly used.

03 There aren't enough endoscopes, so let the CT scanners pick the patients first.

Wang Qifeng clearly stated EAGLE's positioning: "AI is not meant to replace endoscopy, but to screen out high-risk individuals for endoscopic examination, which will make endoscopic examination more accurate and significantly improve the detection rate."

The research team validated EAGLE in high-risk population screening and routine hospital CT scans.

First, consider high-risk groups. Patients undergo plain CT scans, and EAGLE is run in high-sensitivity mode. Those testing positive are prioritized for endoscopic examination.

In July 2023, the team conducted a simulation study in Suining, Sichuan, on "CT initial screening + endoscopic diagnosis". The results showed that the detection rate of malignant lesions by endoscopy increased from 1.7% to 5.2%, and the number of endoscopy procedures required to detect one malignant lesion decreased from 59 to 19.

Different uses of EAGLE in routine hospital CT scans and gastroscopy screening for high-risk populations. Redrawn based on data from a Nature Medicine paper.

This is not a complete prospective clinical outcome. The three esophageal cancer cases in the mixed cohort were from retrospective case resampling; how much endoscopy can reduce esophageal cancer needs to be validated in real-world screening programs.

Another set of validations utilized existing hospital chest CT scans. Lung cancer screening already uses low-dose chest CT scans, and many patients in hospitals also have chest CT scans performed for pneumonia, coughs, and hospitalization examinations. EAGLE searches for esophageal abnormalities on these images without requiring an additional scan; after the AI indicates a positive result, the doctor then determines whether an endoscopy is necessary.

The pneumonia CT scan that Wang Qifeng encountered two years ago belongs to this category.

The requirements for models differ between high-risk screening and routine hospital CT scans. The former prioritizes minimizing missed diagnoses, while the latter, which deals with the majority of individuals without esophageal cancer, requires fewer false positives, thus reducing the burden on doctors for review and patients for follow-up examinations. Wang Qifeng stated that the model's thresholds can be adjusted according to the specific scenario.

For high-incidence areas, the CT scanners already widely available at the grassroots level can handle the initial screening, so there is no need to build a separate imaging facility.

04 The path of plain CT scans leads to the fourth type of cancer.

Since 2021, DAMO Academy has been researching "plain CT scan + AI" for multi-cancer screening, and has successively launched models for pancreatic cancer, gastric cancer, colorectal cancer, and esophageal cancer, also covering major diseases such as aortic dissection. According to DAMO Academy, related research has been published in *Nature Medicine* five times. Among them, the pancreatic cancer model PANDA received FDA "Breakthrough Device" designation in April 2025, and this technical approach has also been adopted by the National Science and Technology Major Project for the "Four Major Chronic Diseases".

A list of medical AI models from DAMO Academy (Image source: DAMO Academy)

The EAGLE project was launched in 2022. Wang Qifeng was willing to participate because of PANDA. His idea at the time was very simple: he treated late-stage patients every day, the treatment process was painful, and the survival period was not long. If he could train a model to detect early lesions from the CT scans he took every day, he could save more lives.

Early cancer screening is still being researched using liquid biopsy. Taking Grail's Galleri as an example, patients need to have their blood drawn separately, and then cancer clues are searched for in cfDNA methylation signals. EAGLE analyzes CT scans already produced by hospitals, which can save an additional initial screening test; a definitive diagnosis still requires subsequent specialist procedures.

Chest CT and abdominal CT cover different organs. Even after an AI (artificial infiltration) test indicates a positive result, the patient still needs to undergo endoscopy, pathology, or other specialist examinations.

Hospitals can reuse existing images, but they also need to undertake more review, referral, and follow-up work.

According to the DAMO Academy's vision, the hospital's existing chest or abdominal plain CT scans can be expanded to include lung cancer, liver cancer, colorectal cancer, stomach cancer, esophageal cancer, pancreatic cancer, and breast cancer. Since the scanning range differs, a single CT scan cannot simultaneously cover all seven types of cancer, thus reusing the same imaging equipment.

EAGLE is still a long way from widespread use. The model's sensitivity for high-grade intraepithelial neoplasia is 52.5%, meaning that nearly half of the lesions were not detected.

In a prospective hospital validation study, 42 out of 90 individuals who tested positive via AI did not proceed with the original treatment process. A multidisciplinary team recommended that 16 of them undergo gastroscopy, but only 3 ultimately completed the procedure, and one case of esophageal cancer was discovered. After an AI report indicates a positive result, the hospital must arrange for imaging review, specialist referral, gastroscopy, and ongoing follow-up.

The completion rate of gastroscopy, the number of newly detected early cases, the cost of false positive retesting, and whether it can reduce esophagectomy, radiotherapy and chemotherapy and late-stage treatment still need to be answered in larger-scale prospective studies.

EAGLE has validated a viable approach: using routine hospital checkups to perform plain CT scans to screen for high-risk individuals with solid tumors, in addition to existing diagnostic and treatment goals.

This article is from the WeChat official account "Hard AI" . For more cutting-edge AI information, please go here.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.