Most AI pathology tools need thousands of labeled tissue images and institution-specific retraining before they can reliably spot a single cancer type. A team led by Prof. Li Xiaomeng at the Hong Kong University of Science and Technology, working with Guangdong Provincial People’s Hospital and Harvard Medical School, says it has built something fundamentally different: a “plug-and-play” system called PRET, for Pan-cancer Recognition without Example Training, that can recognize 18 distinct cancer types from as few as one to eight annotated tumor slides, with no additional training required. The work was published in Nature Cancer on April 21, 2026.
How PRET is different from existing pathology AI
Conventional pathology algorithms are trained on massive labeled datasets specific to one cancer type and one institution’s scanning equipment and staining protocols, meaning a model built for breast cancer at one hospital often has to be retrained from scratch to work at a different hospital or on a different cancer entirely. PRET borrows a concept from large language models called in-context learning, allowing it to adapt to a brand-new cancer type during inference simply by being shown a handful of example slides, rather than undergoing a fresh, resource-intensive training cycle each time.
The performance numbers
According to the published results, PRET was validated against 23 international benchmark datasets and outperformed 20 existing methods, achieving an area-under-curve, or AUC, above 97% in 15 of 20 tasks tested. In colorectal cancer screening it reached 100% AUC, and in esophageal squamous cell carcinoma segmentation it hit 99.54% AUC. Perhaps most striking, in lymph node metastasis detection using just eight reference slides, PRET achieved 98.71% AUC, surpassing a panel of 11 human pathologists whose average accuracy was 81%.
Why this matters for hospitals without specialist pathologists
“The core value of the PRET system lies in breaking down the traditional barriers of ‘massive data and repetitive training,’” Prof. Li Xiaomeng said, according to coverage of the study. In practice, that could mean smaller hospitals or clinics in regions with a shortage of subspecialty pathologists could deploy a single, adaptable model rather than needing separate validated tools for each cancer type, potentially narrowing the diagnostic gap between well-resourced academic medical centers and community or rural hospitals.
Caveats before clinical use
The Nature Cancer paper itself focuses on future development plans rather than cataloguing current shortcomings, and independent researchers reviewing this class of technology typically note that benchmark performance on curated datasets does not guarantee equivalent accuracy in messy, real-world clinical workflows, where slide quality, staining variability, and rare tumor subtypes can all degrade AI performance. The system has not yet been validated in a live clinical environment or gone through a regulatory clearance process comparable to the FDA’s 510(k) pathway used for other AI pathology tools.
Where PRET fits among other pathology AI efforts
PRET arrives as digital pathology overall is drawing heavy investment and research attention. Companies including Paige.AI have built some of the earliest commercial AI pathology models for prostate and breast cancer detection, while venture-backed startups such as Aiforia in Finland and Cellens in Boston are pursuing narrower, single-cancer-type diagnostic tools using deep learning and biophysical measurement respectively. What sets PRET apart from most of these commercial efforts, at least on paper, is its breadth: rather than being built and validated for one cancer type at a time, it is designed from the outset to generalize across 18 of them using the same underlying model, a scope that, if it holds up in prospective testing, would put it in a different category from most FDA-cleared pathology tools currently on the market.
What’s next
The research team’s stated next step is testing PRET prospectively in real clinical settings rather than only against retrospective benchmark datasets. If those results hold up, the technology could represent a meaningful shift for digital pathology, moving the field away from a model of one narrow AI tool per cancer type toward a single adaptable system that hospitals could point at a new tumor type as soon as a handful of confirmed cases are available, potentially compressing years of tool-specific validation into a much shorter runway.
