AI Cancer Detection Spots Hidden Tumors as FDA Fast-Tracks PANDA

The breakthrough is AI seeing what people miss

This month's cancer AI news doesn't come down to one discovery. What the reporting describes is a pattern across several studies: models are finding tumors and treatment signals in medical images and tissue samples that look normal to trained specialists. Three developments from roughly the same week show it. A pancreatic cancer detector received a federal regulatory designation. A breast cancer model passed a test against randomized trial data. And a run of coverage of imaging AI argues that the most important cancers may be the ones nobody can see yet.

Taken together, the sources point to one conclusion. AI is having its biggest effect in oncology where human perception runs out, not where it helps people work faster. The same sources also suggest that the step from a strong validation study to a real change in survival is longer than the headlines imply.

What happened: PANDA gets an FDA breakthrough designation

The clearest regulatory news concerns DAMO PANDA, a model built by DAMO Academy to find pancreatic cancer lesions on CT scans. The FDA has given it breakthrough device designation, which speeds up review of the device.5 The supporting evidence comes from a real-world validation in 20,530 patients. PANDA reached 92.9% sensitivity and 99.9% specificity for pancreatic ductal adenocarcinoma. That beat average radiologist performance by 34.1 points on sensitivity and 6.3 points on specificity.5

The study authors' explanation of why the model works is the most interesting part for AI research. They say deep learning systems respond more strongly to small differences in grayscale intensity than human eyes do, because people are better at noticing changes in color than in brightness.5 That is a specific claim that machines have a perceptual advantage, not just more speed or stamina. The researchers were also able to find lesions on non-contrast CT, which uses less radiation, avoids reactions to contrast agents, and costs less.5 That point matters for screening large populations.

PANDA was trained on 3,208 patients, launched in November 2023, and has been used in Chinese hospitals and routine health checks. Its developers say the approach could extend to lung, breast and liver cancers, and DAMO has partnered with the World Health Organization to bring the tool to low-resource regions.5

The Mayo Clinic work points the same way

PANDA is not working alone. A feature on AI cancer screening describes REDMOD, a radiomics-based early-detection model built at Mayo Clinic by radiologist Ajit Goenka.6 It was trained on CT scans taken from patients 3 to 36 months before their pancreatic cancer diagnosis, alongside control scans covering about 15 years of practice. It reads small patterns in the brightness and texture of the pancreas. The coverage says it can flag signs of cancer about 16 months earlier than specialist radiologists and at almost twice their rate.6

Mayo's own account fills in the background. Earlier screening programs for high-risk patients mostly failed because people reviewing scans could not reliably spot disease that early, and false-positive rates climbed.1 What the institution had was decades of longitudinal imaging. Researchers could look back at scans taken a year or more before cancer was found and train a model on them. Mayo's medical director for generative AI put its accuracy at 97%.1 The algorithm is now being tested in a clinical trial called AI-PACED, which focuses on high-risk patients, especially those with a family history of the disease.1

The problem these tools target is well documented. Pancreatic cancer's five-year relative survival rate is about 13%.51 In the same feature, gastroenterologist Michael Wallace reviewed CT and MRI scans taken one month to three years before diagnosis. He found that up to half showed a pancreatic abnormality that could have led to more testing.6 In other words, the signal was often in the images already.

Pathology slides as a source of hidden information

The second development moves from radiology to the microscope. At the Technion, researchers with partners in Israel, the US, Europe and Australia built a deep learning model that reads ordinary hematoxylin-and-eosin-stained pathology slides. It predicts both whether early breast cancer will come back and whether a patient will benefit from chemotherapy.4 The study was published in The Lancet Oncology and presented at the European Society for Medical Oncology congress. It is described as the first AI model of its kind validated against a large randomized clinical trial.4

The research design deserves attention. The team did not run a new trial. They applied the model to archived slides from TAILORx, a breast cancer study of more than 10,000 patients in which chemotherapy was assigned at random.4 Because of that randomization, the researchers could measure whether chemotherapy actually helped the patients the model identified, not just whether those patients did well. According to the reporting, the model found patients who benefit from chemotherapy and large groups, especially postmenopausal women, who do not. It performed about as well as widely used genomic tests and held up across hospitals on four continents.4

This matches a broader trend described elsewhere. Models trained on large collections of digitized slides can now infer molecular information from cheap stains that almost every biopsy already gets.7 One example is EAGLE, built at Memorial Sloan Kettering and Mount Sinai. It flags EGFR mutations in lung cancer and could reduce the need for rapid genetic tests by up to 43%.7

Where the coverage agrees, and where it doesn't

The coverage agrees that AI can pull clinically useful signal out of data humans already collect, including CT scans, mammograms and stained tissue. Wider reviews make the same case: radiology models extract high-dimensional features that the eye cannot measure, and multimodal systems that combine data types do better than single-source models.103 A Hong Kong framework called PRET reportedly reached 98% accuracy on lymph-node metastasis detection and outperformed 11 pathologists.6

The coverage splits on how much this changes outcomes. A physician writing in Forbes describes AI's real contribution as a set of incremental gains, not one big cure.7 His evidence is useful. In Sweden's MASAI trial of 105,934 women, AI-assisted mammography raised cancer detection from 73.8% to 80.5% with no increase in false alarms. But a commentary in the same journal calculated that per 1,000 women screened, AI found 1.4 extra cancers and prevented only 0.2 of the cancers that show up between screenings.7 Finding more cancers is not the same as preventing more deaths. Overdiagnosis is a known risk whenever detection becomes more sensitive.

Generalization is the second weak point. EAGLE held up across American and European patients. However, a 2026 JAMA Oncology analysis found its accuracy fell to about 0.68 in patients of Asian ancestry and 0.66 in samples from the lining of the lung.7 That is why the multi-continent validation in the Technion study and the external-center results reported for PANDA matter so much.45

Design as well as detection

One research line goes past detection to building treatments. A team at the Technical University of Denmark and Scripps Research reported in Science that it could design protein "minibinders" entirely on a computer. These guide T cells to cancer cells by recognizing pMHC molecules, and the method cut the time to a lead molecule from years to four to six weeks.2 The researchers also ran AI screens against pMHC molecules on healthy cells to remove designs likely to cause dangerous cross-reactions before lab testing.2 The lead researcher still expects up to five years before human trials begin.2 Even when the computation is fast, the biology sets the pace.

What it means

The most defensible reading is that cancer AI has moved from automating what clinicians do to perceiving what they can't. The breakthrough designation for PANDA, the randomized-data validation of the Technion model, and REDMOD's lead time of more than a year all support that view.546 The next phase will depend on prospective trials such as AI-PACED, and on whether earlier detection and better-targeted treatment actually reduce deaths, not just raise detection counts.17 For AI research, that is a harder benchmark than accuracy on a held-out dataset, and it is the one that counts.