AI Cancer Screening: DAMO Academy's PANDA Detects Pancreatic Cancer from Non-Contrast CT, FaceAge Predicts Survival Rates Using Photos

The system scanned that non-contrast CT and flagged a lesion about 1.5 centimeters in diameter in the pancreatic region, suggesting a possible malignancy. The attending physician was taken aback—Mr. Qiu's symptom was dizziness, with nothing pointing to the pancreas.

Three days of dizziness, and an unexpected encounter with the "king of cancers"

Zhu Kelei, vice president of Ningbo University Affiliated People's Hospital and chief physician of the Department of Hepatobiliary and Pancreatic General Surgery, later recalled: "This is something none of us doctors could have imagined, because the patient's chief complaint was dizziness, and pancreatic problems usually present as abdominal pain."

The AI alert prompted the hospital to quickly adjust course, and a contrast-enhanced CT was arranged immediately, with the results clearly pointing to pancreatic cancer. Mr. Qiu soon underwent surgery, and the pathology report confirmed pancreatic ductal adenocarcinoma. His daughter said: "We're so thankful—my father truly got a second chance at life."

Pancreatic cancer has long been known as the "king of cancers." The pancreas lies deep in the abdominal cavity, surrounded by the stomach, duodenum, and spleen; its location is hidden, and early-stage pancreatic cancer has almost no specific symptoms. By the time patients feel obvious discomfort, the cancer cells have often already quietly metastasized.

China has more than 130,000 new pancreatic cancer patients each year, and more than 80% are diagnosed at an advanced stage, with a five-year survival rate of less than 10%.

Starting in November 2024, Ningbo University Affiliated People's Hospital introduced the medical AI of Alibaba DAMO Academy (Hupan Lab) and, together with Professor Jin Gang's team at the Shanghai Pancreatic Disease Research Institute, launched a prospective clinical trial, becoming the first hospital in Zhejiang Province to open a pancreatic cancer early screening clinic.

To date, the Ningbo project has found 8 cases of pancreatic cancer from the CT imaging data of more than 40,000 people, 6 of which were early-stage, and all of them underwent surgical treatment.

The DAMO Academy PANDA model research team registered the contrast-enhanced CT images of more than 2,000 pancreatic cancer patients onto non-contrast CT scans, teaching the model to identify subtle pancreatic lesions in lower-resolution images.

The model's sensitivity for identifying pancreatic cancer patients reached 92.9%, with specificity of 99.9%, about 34% higher sensitivity than experienced radiologists.

Contrast-enhanced CT requires injection of a contrast agent and is relatively expensive, making it unsuitable for large-scale population screening; non-contrast CT, by contrast, is the most common imaging method in outpatient settings. If "opportunistic screening" for cancer could be completed within these readily available images, it would mean no additional examination burden for patients, and any ordinary visit or health checkup could become an opportunity to detect early cancer.

DAMO Academy's multi-cancer early screening AI technology has begun global rollout, covering 9 countries and regions worldwide and serving more than 20 million people.

At the same time, DAMO Academy is also advancing the goal of "one scan, multiple screenings"—that is, through a single non-contrast CT, AI can not only identify pancreatic cancer but also simultaneously screen for liver cancer, esophageal cancer, gastric cancer, colorectal cancer, and even emergencies such as aortic syndrome.

Following pancreatic cancer, DAMO Academy's colorectal cancer screening AI model DAMO COCA accurately identified 5 missed cases of colorectal cancer from the non-contrast CT images of 27,000 people, achieving sensitivity of 86.6% and specificity of 99.8%.

Doctors are irreplaceable in experiential judgment and clinical decision-making, but the human eye has fatigue blind spots, attention is drawn by the chief complaint, and when facing hundreds of detail-dense image slices, there is always the possibility of omission.

Recently, a research team at Harvard University's Massachusetts General Hospital published a study in The Lancet Digital Health in which they used an ordinary photograph to predict the mortality risk of a cancer patient. Although the model's predictions are not entirely accurate, those patients with higher mortality risk scores did indeed have a higher probability of death.

FaceAge's logic is built on the concept of "biological age." Two people both 60 years old may look 52 or 71 because of differences in lifestyle, disease history, and genetics.

Researcher Mak noted that people who smoke often look older than their actual age; the same is true for heavy drinkers.

The researchers trained an algorithm to analyze and process 40 million images of ordinary faces without being given a specific task. They then fine-tuned the model so that it could analyze facial images in two different ways: one for estimating a person's physiological age, and one for predicting the survival odds of cancer patients.

Whether the person in the photo is frowning or smiling, whether they are wearing makeup, and what lighting conditions they are in—none of these affect the age prediction.

For mortality risk prediction, in a test group of 3,820 cancer patients, more than half of those in the group with the highest mortality risk scores died within just 90 days after the photo was taken.

The researchers also found in the genetic data of lung cancer patients that FaceAge scores were negatively correlated with a gene called CDK6. CDK6 is an important participant in cell cycle regulation and is associated with the progression of multiple cancers. A model trained on facial photos had, to some extent, captured signals at the genetic level.

Sophont's CEO shared another case on social media, involving osteosarcoma.

Their model Holon-S1 analyzed the patient's pathology slides. This model had not undergone specialized training for osteosarcoma, but it extracted key biological information from the slide images: the patient's fibroblast-related genes showed overexpression.

This phenomenon had previously been discovered by the patient's medical team through advanced molecular testing, and the corresponding therapeutic target had already been identified.

New questions

In their paper, the FaceAge researchers specifically mentioned a direction that worries them: such a tool may be a help to clinicians, but it may also be highly attractive to life insurance companies or employers, who might try to use it to assess an individual's risk level.

In addition, the "visual features of aging" learned by the model may well not apply to all skin colors and facial structures. If the model systematically overestimates biological age for certain ethnic groups, this error would replicate and amplify existing inequality in an extremely subtle way, and those misjudged might not even know what happened.

Previous research has found that some commercial models used to predict patients' medical needs systematically underestimate the needs of Black patients, because the models use historical medical spending as a proxy for health status, while historically Black patients have received fewer medical resources for the same condition.

The researchers emphasized that before FaceAge enters clinical use, it needs to be tested in more diverse populations to identify and eliminate potential errors.

How to handle these new questions is far more difficult than training a model.

This phenomenon had previously been discovered by the patient's medical team through advanced molecular testing, and the corresponding therapeutic target had already been identified.

New questions

In their paper, the FaceAge researchers specifically mentioned a direction that worries them: such a tool may be a help to clinicians, but it may also be highly attractive to life insurance companies or employers, who might try to use it to assess an individual's risk level.

In addition, the "visual features of aging" learned by the model may well not apply to all skin colors and facial structures. If the model systematically overestimates biological age for certain ethnic groups, this error would replicate and amplify existing inequality in an extremely subtle way, and those misjudged might not even know what happened.

Previous research has found that some commercial models used to predict patients' medical needs systematically underestimate the needs of Black patients, because the models use historical medical spending as a proxy for health status, while historically Black patients have received fewer medical resources for the same condition.

The researchers emphasized that before FaceAge enters clinical use, it needs to be tested in more diverse populations to identify and eliminate potential errors.

How to handle these new questions is far more difficult than training a model.