A “plug-and-play” AI recognizes 18 cancer types from just a handful of slides

Most medical AI tools share a stubborn limitation: they are trained for a specific task at a specific hospital, and moving them to a new institution, imaging system or patient population usually means retraining them from scratch. That fragility is one reason powerful diagnostic models rarely reach the clinics that need them most. A system described in Nature Cancer in April 2026 takes a different approach.

Called PRET, the system was built by a research team led by Prof. Li Xiaomeng at the Hong Kong University of Science and Technology, in collaboration with Guangdong Provincial People’s Hospital and Harvard Medical School. Its central idea is borrowed from the language models that power modern chatbots: “in-context learning.” Instead of being fine-tuned for each new task, PRET adapts on the fly by looking at just one to eight annotated example slides for a cancer type it has not seen before — much as a person can learn a new category from a handful of examples.

“That fragility is one reason powerful diagnostic models rarely reach the clinics that need them most.”

The performance held up across a demanding evaluation spanning 18 cancer types and 23 international benchmark datasets from institutions in mainland China, the United States and the Netherlands. PRET’s area under the curve — a standard measure of diagnostic accuracy — exceeded 97% on 15 tasks, reaching 100% in colorectal cancer screening and 99.54% in segmenting esophageal squamous cell carcinoma. In detecting lymph node metastasis, it achieved an AUC of about 98.71% using only eight reference slides, surpassing the average of 11 pathologists, whose AUC averaged roughly 81%.

The practical value lies precisely in that flexibility. A single adaptable system that a hospital can point at a new cancer type without a costly retraining project could put expert-level pathology support within reach of smaller labs and regions with few specialist pathologists. Rather than replacing pathologists, such a tool could handle routine screening and flag hard cases for human review, freeing experts to focus where their judgment matters most.

As with any benchmark result, the leap to everyday diagnosis will require prospective validation in real clinical workflows and careful oversight to ensure the tool performs fairly across different populations and tissue-preparation methods. But the design philosophy is quietly important: an AI that learns from a few examples, rather than demanding a mountain of institution-specific data, is exactly the kind of tool that could travel to where care is scarcest.