What is AI pancreatic cancer detection on CT?
AI pancreatic cancer detection uses deep learning software to look for tumours and other lesions in the pancreas on CT scans. The best-known example is PANDA, short for pancreatic cancer detection with artificial intelligence, developed by researchers in China with Alibaba's DAMO Academy and published in Nature Medicine in 2023.
What makes PANDA different is that it reads non-contrast CT, the plain scan done without dye injected into a vein. Non-contrast CT is often done for other reasons, such as chest or abdominal checks, and the pancreas is usually hard to judge on it. The idea is that the AI could check these scans in the background and flag people who may need proper tests.
This page explains how PANDA was built, what it showed in hospital test sets and in a large real-world group of 20,530 patients, what later use in China has reported, the limits of the evidence and what it may mean for patients in India.
Key facts about the PANDA AI pancreatic cancer detection study
The main published evidence, in brief:
- Publication: Cao K, et al. Nature Medicine, November 2023; authors from China, with Alibaba DAMO Academy
- Training: 3,208 patients from one centre in Shanghai, 2,270 with pancreatic lesions and 938 without
- Internal test (291 patients): AUC 0.996 for finding any lesion; 92.6% sensitivity and 97.3% specificity for pancreatic cancer (PDAC)
- External test (5,337 patients, nine centres in mainland China, Taiwan and the Czech Republic): 93.3% sensitivity and 98.8% specificity for any lesion
- Reader study: PANDA's sensitivity for PDAC was 34.1% higher than the average of 33 radiologists reading the same plain CT scans
- Real-world test: 20,530 consecutive patients, reported in the abstract as 92.9% sensitivity and 99.9% specificity for lesion detection
- Regulatory: the company said in April 2025 that the US FDA granted breakthrough device designation, which is not an approval
Key takeaways: AI pancreatic cancer detection and PANDA
- PANDA found pancreatic lesions on plain CT scans with high accuracy in retrospective hospital data from China and two other sites.
- It did better than the average radiologist reading the same plain scans, and helped radiologists when they used it.
- In real-world data it raised few false alarms per scan, but because pancreatic cancer is uncommon, many flags will still not be cancer.
- Telling one type of pancreatic lesion from another was much less accurate than spotting that a lesion was there.
- There is no published evidence yet that using PANDA helps people live longer; prospective studies are still running.
How does PANDA read a non-contrast CT scan?
PANDA was trained on scans from 3,208 patients at one centre in Shanghai. Of these, 2,270 had a pancreatic lesion and 938 had a normal pancreas. The lesions included 1,431 pancreatic ductal adenocarcinomas (PDAC, the most common pancreatic cancer), as well as neuroendocrine tumours, cysts and chronic pancreatitis.
On plain CT, tumours are hard to outline. So the team drew the lesions on each patient's contrast-enhanced CT, where they are clearer, and then transferred these outlines onto the matching plain CT by lining up the two scans. The AI learned to find the lesion on the plain scan alone.
PANDA does three jobs. It says whether a lesion is likely present, whether it looks like PDAC, and which of eight lesion types it most resembles. Because it reads scans that were already done, a New York Times report from a hospital using it notes there is no extra test cost for the patient. A flag is a reason for doctors to look again, not a diagnosis.
What did the PANDA study show in hospital test sets?
In an internal test of 291 patients from the same centre, PANDA's area under the curve (AUC) was 0.996 for detecting any lesion, with 94.9% sensitivity and 100% specificity. For identifying PDAC, the AUC was 0.987, with 92.6% sensitivity and 97.3% specificity. AUC measures how well a test separates people with and without disease; 1.0 is perfect and 0.5 is chance.
The team then tested it on 5,337 patients from nine other centres: seven in mainland China, one in Taiwan and one in Prague in the Czech Republic. Here PANDA detected lesions with an AUC of 0.984, 93.3% sensitivity and 98.8% specificity. For PDAC it had 90.1% sensitivity and 95.7% specificity. The paper's abstract describes the overall multicentre validation as 6,239 patients across 10 centres, with AUCs from 0.986 to 0.996.
These were retrospective tests on scans collected before surgery, in groups with far more cancers than in the general population. That makes accuracy look different from everyday screening.
How did PANDA compare with radiologists?
In a reader study, 33 radiologists from 12 institutions read the same 291 plain CT scans. Compared with the average reader, PANDA's sensitivity for identifying PDAC was 34.1% higher and its specificity 6.3% higher. For detecting any lesion, it was 14.7% higher in sensitivity and 6.8% in specificity. When radiologists read the scans again with PANDA's help, their sensitivity for PDAC rose by 20.5% and specificity by 3.1%.
A second comparison was harder for the AI. Fifteen pancreatic specialists read contrast-enhanced CT, the scan normally used to look for pancreatic cancer. PANDA, using only the plain scan, still had 13.0% higher sensitivity and 0.5% higher specificity for PDAC than these specialists.
These comparisons show what the software can do on a fixed set of scans in a study. They do not show how it would perform alongside radiologists in daily work, where many scans are normal and doctors have other clinical information.
PANDA study numbers at a glance
Reported results from the Nature Medicine paper:
- Internal test (291): lesion AUC 0.996; PDAC sensitivity 92.6%, specificity 97.3%
- External test (5,337, nine centres): lesion sensitivity 93.3%, specificity 98.8%; PDAC sensitivity 90.1%, specificity 95.7%
- Readers (33 radiologists): PANDA 34.1% higher sensitivity and 6.3% higher specificity for PDAC
- Lesion type (eight classes): accuracy 79.6% internal and 81.4% external
- Real-world first group (16,420): PDAC sensitivity 95.5% (44 PDACs); specificity 99.8%
- Real-world positive predictive value for PDAC: 56.0% (68.9% after adjustment)
What happened when PANDA was tested on 20,530 real-world patients?
The team ran PANDA on 20,530 consecutive patients who had CT scans for many reasons, including health check-ups, emergency visits, outpatient care and hospital stays. The paper's abstract summarises the result as 92.9% sensitivity and 99.9% specificity for lesion detection. This was still a retrospective study: the AI read scans after the fact.
The detailed results give a fuller picture. In the first and larger group of 16,420 patients, there were 44 PDACs and 135 other pancreatic lesions. PANDA found 95.5% of the PDACs. Across all lesion types, it found 84.6%, and most of the lesions it missed (89%) were benign cysts. It wrongly flagged 0.5% of patients, 76 people. Of every 100 people it flagged as possible PDAC, 56 had it (about 69 after the authors' adjustment).
In the second group of 4,110 patients, an updated version called PANDA Plus kept the same sensitivity and cut false positives by more than 80%. The authors report that PANDA found 5 cancers and 26 other pancreatic lesions that had been missed in routine care.
Can AI tell types of pancreatic lesions apart?
Finding that something is there is only the first step. Doctors also need to know what it is, because a harmless cyst, a cyst that needs watching, a neuroendocrine tumour and PDAC are managed very differently.
PANDA tried to sort lesions into eight classes: PDAC and seven other types. Overall accuracy was 79.6% in the internal test and 81.4% in the external test, and the authors report this was not worse than radiology reports based on contrast-enhanced CT. But balanced accuracy, which gives rarer lesion types equal weight, was only 60.7% and 52.6%. That means the AI was much less reliable for the less common lesion types.
The authors themselves note relatively low accuracy for pancreatic neuroendocrine tumours. In practice, a lesion flagged on plain CT still needs proper imaging, such as contrast-enhanced CT or other tests chosen by the doctor, and often a specialist opinion.
False alarms in AI pancreatic cancer detection
Pancreatic cancer is uncommon in the general population. That makes false alarms a real problem, even for a very specific test. The PANDA authors estimate that if the tool were used to screen people where PDAC affects 13 in 100,000, only about 1 in 10 people flagged for PDAC would actually have it. This is a model calculation, not a measured result.
Real-world use shows the same issue. The New York Times reported in January 2026 that at a hospital in Ningbo, China, about 1,400 scans had triggered alerts, and about 300 needed follow-up. Doctors there said the tool sometimes flags pancreatitis, inflammation of the pancreas, and cannot tell whether a tumour began in the pancreas or spread there from elsewhere.
A false alarm can mean worry, more scans and sometimes procedures. That is one reason doctors are careful about screening people with no symptoms for pancreatic cancer.
Has PANDA been used in real hospitals since the study?
Yes, in China, but published results are limited. According to a New York Times report from January 2026, the Affiliated People's Hospital of Ningbo University began using PANDA in November 2024 as part of a clinical trial. By then it had analysed more than 180,000 abdominal or chest CT scans and helped find about two dozen pancreatic cancers, 14 of them at an early stage. These are figures reported to journalists, not peer-reviewed results.
A related prospective study is registered on ClinicalTrials.gov (NCT06638866). It tests a deep learning system built with technical support from Alibaba DAMO Academy (the record does not use the name PANDA), with Changhai Hospital in Shanghai as lead sponsor and other Chinese hospitals including Yinzhou Hospital in Ningbo. It aims to enrol about 5,000 people having routine plain CT for other reasons, with primary completion expected on 31 December 2029.
In April 2025, Alibaba's DAMO Academy said the US FDA had granted PANDA breakthrough device designation. That status is meant to speed up review. It is not an approval, and we found no report of US marketing authorisation.
Limits of the AI pancreatic cancer detection evidence
What the PANDA evidence cannot yet tell us:
- All published accuracy results are retrospective; the AI read scans after the fact.
- Most data came from Chinese hospitals; the authors note limited data outside East Asian populations.
- Test groups had many more cancers than the general population, which flatters accuracy.
- Accuracy for rarer lesion types, especially neuroendocrine tumours, was much lower.
- There is no evidence yet that using PANDA leads to longer survival, and cost-effectiveness has not been assessed.
- Real-world figures from Ningbo come from news reports, not a peer-reviewed publication.
Who is pancreatic cancer screening meant for?
Routine screening for pancreatic cancer is not recommended for most people. In 2019, the US Preventive Services Task Force recommended against screening adults who have no symptoms and are not known to be at high risk. It found no evidence that screening reduces deaths, and judged the harms, including false positives and treatment risks, to be at least moderate.
PANDA's developers see it as a different idea, often called opportunistic screening. The AI would check plain CT scans that people are already having for other reasons, so no extra scan is needed. The authors say it is meant as a first step before diagnosis, not a replacement for diagnostic imaging.
Whether this approach does more good than harm in everyday care has not been shown yet. People with a strong family history or known inherited risk may have a different plan, and that should be discussed with a specialist.
AI pancreatic cancer detection evidence by region
Where the evidence comes from, in brief:
- China: PANDA training data, most test centres, the 20,530-patient real-world test and the Ningbo hospital use
- Taiwan and Czech Republic: one external test centre each in the 2023 study
- United States: breakthrough device designation reported by the company in April 2025; no approval reported
- India: no published PANDA test in Indian patients found
Is AI pancreatic cancer detection available in India?
We did not find any published study testing PANDA on Indian patients, or any report of it being approved or routinely used in India. The study authors note that their data came mostly from East Asian populations and hospitals, so its accuracy in Indian patients, on Indian scanners and in Indian hospital workflows is not known.
This does not mean Indian patients cannot benefit from careful CT reading. If a CT scan done for another reason mentions something in the pancreas, the usual path is for the doctor to decide on further imaging and, if needed, referral to a specialist team. An AI flag would be one more input into that process, not a diagnosis.
Families considering treatment or second opinions in other countries, including China, may hear about AI tools. It is reasonable to ask which tool is used, what evidence supports it, and who reviews its results. Your doctor can explain whether any of this applies to you.
What happens next for AI pancreatic cancer detection research?
The PANDA authors say further validation is needed in external real-world centres, in international groups and in prospective studies, and that cost-effectiveness and the balance of benefits and risks still have to be assessed.
The related prospective study NCT06638866 in China, which started in August 2024, will measure how many PDACs and high-risk precursor lesions are found in people having routine plain CT, how many people are recalled, and how many flags turn out to be correct. Its primary completion is expected at the end of 2029. The Ningbo hospital's experience, reported in the media, has not yet been published as a peer-reviewed study.
The key questions are whether finding pancreatic cancer earlier with AI leads to more people having curative surgery and living longer, and whether the extra follow-up tests are worth it. Those answers will take several years.
Questions to ask your doctor about AI pancreatic cancer detection
- My CT report mentions the pancreas. What does this finding mean, and how sure are we?
- Was any AI software used to read my scan, and which one?
- Do I need a contrast-enhanced CT, MRI or another test to look more closely?
- Should I see a pancreas specialist or have my case reviewed by a multidisciplinary team?
- Given my family history or other conditions, am I at higher risk of pancreatic cancer?
- If the finding is a cyst, how will it be followed up, and how often?
Sources
1. Cao K, et al. Large-scale pancreatic cancer detection via non-contrast CT and deep learning. Nature Medicine, 2023. https://doi.org/10.1038/s41591-023-02640-w 2. Cao K, et al. Large-scale pancreatic cancer detection via non-contrast CT and deep learning (open-access full text). Europe PMC, 2023. https://europepmc.org/article/PMC/PMC10719100 3. Alshybani I. Commentary on 'Large-Scale Pancreatic Cancer Detection via Non-Contrast CT and Deep Learning'. Biomedical Engineering and Computational Biology, 2024. https://doi.org/10.1177/11795972241293521 4. Targeted Oncology. AI tool earns FDA breakthrough device designation in pancreatic cancer. Targeted Oncology, 2025. https://www.targetedonc.com/view/ai-tool-earns-fda-breakthrough-device-designation-in-pancreatic-cancer 5. The New York Times (republished by The Nightly). AI technology PANDA is helping doctors find deadly tumours in China. The Nightly, 2026. https://thenightly.com.au/from/the-new-york-times/the-new-york-times-ai-technology-panda-is-helping-doctors-find-deadly-tumours-in-china-c-21186581 6. ClinicalTrials.gov. AI-powered Early Detection for Pancreatic Cancer Via Non-contrast CT in Opportunistic Screening Cohort (NCT06638866). ClinicalTrials.gov, 2025. https://clinicaltrials.gov/study/NCT06638866 7. US Preventive Services Task Force. Pancreatic cancer: screening. USPSTF, 2019. https://www.uspreventiveservicestaskforce.org/uspstf/recommendation/pancreatic-cancer-screening