A pathology AI system developed at the Hong Kong University of Science and Technology can recognize 18 different cancer types after seeing as few as one to eight annotated tissue slides, without needing to be retrained for the hospital or patient population where it’s deployed, according to research published in Nature Cancer in 2026.
The system, called PRET, for Pan-cancer Recognition without Example Training, was developed by a team led by Prof. Li Xiaomeng at HKUST in collaboration with Guangdong Provincial People’s Hospital and Harvard Medical School, according to HKUST’s announcement and reporting from Medical Xpress and MobiHealthNews.
Borrowing an idea from language models
PRET’s key technical innovation is applying “in-context learning” — a concept familiar from large language models, where a model adapts to a new task just by being shown a handful of examples in its prompt, without retraining — to pathology image analysis. In practice, that means a pathologist can hand PRET a small number of annotated slides for a cancer type or task it hasn’t specifically been trained on, and the model adapts on the spot during inference, performing tasks like cancer screening, tumor subtyping, and tumor segmentation.
Why “no retraining” is the actual breakthrough
Most AI pathology tools to date have needed to be fine-tuned on data from each new institution, cancer type, or scanner setup before they perform reliably — an expensive, slow process that has limited how widely those tools can spread beyond the hospitals that built them. A system that works out of the box across different institutions, scanners, and patient populations addresses one of the biggest practical barriers to scaling pathology AI globally, particularly in settings that lack the data volume or technical staff to train a custom model.
How it was tested
The research team validated PRET using 23 international benchmark datasets sourced from medical institutions in China, the United States, and the Netherlands, spanning the full set of 18 cancer types and a range of diagnostic tasks. PRET outperformed existing methods on 20 of the tasks tested, and its area under the curve exceeded 97% on 15 of those tasks. In two specific benchmarks, the system reached an AUC of 100% for colorectal cancer screening and 99.54% for esophageal squamous cell carcinoma tumor segmentation. Separate reporting on the system also noted it outperformed pathologists in detecting lymph node metastasis, a notoriously labor-intensive task that involves scanning large tissue sections for small clusters of cancer cells.
Where the skepticism sits
Pathology researchers outside the study are likely to note that benchmark performance on curated international datasets, however diverse, is still a step removed from routine deployment inside a working pathology lab, where slide quality, staining variation, and scanner idiosyncrasies can be messier than in benchmark collections. There is also the broader question facing any diagnostic AI aimed at “plug-and-play” deployment: regulatory bodies typically require evidence specific to the intended clinical use and population before a tool can be used for actual patient care decisions, meaning strong benchmark AUCs are necessary but not sufficient for clinical adoption.
What it could mean for pathology departments
If PRET’s out-of-the-box performance holds up in prospective clinical use, tools like it could meaningfully lower the barrier for smaller hospitals and pathology labs — especially in regions without the resources to build institution-specific AI models — to access multi-cancer AI support. Combined with a global shortage of pathologists in many countries, a genuinely generalizable pathology AI could extend specialist-level triage and second-opinion support to labs that currently have no access to it at all. The HKUST team’s next step, consistent with how most diagnostic AI research proceeds after a strong benchmark result, would be prospective clinical validation in partnership with hospitals such as Guangdong Provincial People’s Hospital to see whether the in-context learning approach holds up on real, unscreened patient samples processed under normal lab conditions.
A different model than most pathology AI
Most existing pathology AI systems are built as narrow specialists — a model trained to detect breast cancer metastases in lymph nodes, for instance, needs a separate model, and separate training data, to detect a different cancer type or a different tissue. PRET’s pitch is that a single underlying system can be redirected toward a new diagnostic task simply by showing it a small number of labeled examples, the same way a large language model can be redirected toward a new writing task with a handful of examples in its prompt rather than retraining from scratch. That architecture, if it scales, would let a single deployed system take on new diagnostic tasks over time as pathologists feed it new labeled cases, rather than requiring a hospital to procure and validate an entirely new AI product for each cancer type it wants covered.
The collaboration behind PRET — spanning HKUST’s engineering school, a major Chinese public hospital in Guangdong Provincial People’s Hospital, and Harvard Medical School — also reflects how multi-cancer pathology AI research has become an explicitly international effort, with the 23 benchmark datasets used for validation drawn from institutions across China, the United States, and the Netherlands specifically to test whether the model’s performance holds up across different populations, scanners, and staining protocols rather than just one country’s pathology labs.
