Why is pancreatic cancer so dangerous?
Pancreatic ductal adenocarcinoma, the most common form, is the deadliest solid cancer. It is usually found late, when it can no longer be operated on. Yet studies show that early or incidental detection goes with longer survival.
The problem is that you cannot screen the whole population with a single test: the disease is rare, and a test that gets it wrong (false positives) can cause anxiety and needless examinations.
What is PANDA and how was it tested?
A CT scan taken without contrast dye is a very common examination, done for many reasons. Spotting pancreatic cancer on it was considered impossible, because the tumour is very hard to tell apart there. So the team built PANDA, a deep-learning AI (a program that learns from examples) trained on 3,208 patients from a single hospital.
It was then tested in 10 centres on 6,239 patients, and in real-world conditions on 20,530 consecutive patients, across several care settings.
What did the AI manage to do?
For lesion detection, PANDA's quality measured by AUC (a score where 1 means perfect separation) is 0.986 to 0.996 in the multicentre validation. It beats the mean radiologist performance by 34.1% in sensitivity (the share of true cases found) and by 6.3% in specificity (the share of healthy people correctly cleared) for identifying pancreatic cancer.
In the real world, across 20,530 consecutive patients, sensitivity is 92.9% and specificity 99.9% for lesion detection. For telling common lesion subtypes apart, PANDA on non-contrast CT is no worse than radiologists' reports made with a contrast-enhanced scan.
Key figures
- 92.9% / 99.9%sensitivity and specificity of PANDA for detecting a pancreatic lesion, in real-world conditions
- 34.1%more sensitivity than the average radiologist for identifying pancreatic cancer (plus 6.3% specificity)
- 6,239patients in the multicentre validation, across 10 centres
What does this change for patients?
The idea is to use scans already done for other reasons to find lesions nobody was looking for. Because pancreatic cancer is typically detected late, any extra chance to spot it sooner is valuable, and scans that already exist are an attractive route. The authors believe PANDA could serve as a large-scale screening tool, which remains to be demonstrated.
An AI flag is not a diagnosis: it invites further tests, decided by doctors.
What are the limits?
The study assesses the AI's ability to spot lesions, not yet the effect of screening on survival. Even with very high specificity, because the cancer is rare some alerts do not correspond to cancer: in the first real-world evaluation (16,420 patients), of the cases flagged as pancreatic cancer, 56.0% were correct.
Some false alarms involved other diseases around the pancreas that deserved the radiologists' attention. Performance will need to be confirmed in prospective studies, meaning studies that follow patients over time.
Pancreas: asking the right questions with Mima
PANDA remains a research effort: your scan is not read this way in routine care. The idea is to get more from scans already done, which is still to be shown.
Mima can keep your reports and appointments in one place and gather your questions. A concrete question: "What should be monitored for my pancreas, and how often?"
Frequently asked questions
Why a CT scan without contrast dye?
Because it is a common examination, already done for many reasons. Being able to spot pancreatic lesions on it would make wide-scale detection conceivable without an extra exam.
Can PANDA diagnose pancreatic cancer?
No. It finds and classifies lesions, but diagnosis remains medical and needs further tests.
Is this tool already used in hospitals?
The study published in Nature Medicine in 2023 tests it on consecutive patients in real-world conditions, but it does not show an effect on survival. The authors describe a potential screening tool.
What are sensitivity and specificity?
Sensitivity is the share of true cases found (92.9% here for lesions) and specificity the share of healthy people correctly cleared (99.9%).
The original study
Cao K, Xia Y, Yao J, et al. “Large-scale pancreatic cancer detection via non-contrast CT and deep learning.” Nature Medicine, 20 November 2023. DOI : 10.1038/s41591-023-02640-w
Read the full study in Nature MedicineThis article summarises a scientific study and is not medical advice.