The central obstacle to AI in pathology is not detection accuracy, but deployment. When hospitals want to adopt an AI system for diagnosis, they must retrain it on site, adapting it to local scanners, staining methods, and patient populations. That takes months and requires resources many institutions lack. A system called PRET, developed by researchers at Hong Kong University of Science and Technology, addresses this bottleneck and recognizes 18 cancer types without hospital-specific training.
What PRET is and how it works
PRET stands for "Pan-cancer Recognition without Example Training." It was developed by researchers led by Professor Li Xiaomeng at HKUST. The study appeared in Nature Cancer in April 2026, one of the most prestigious journals in cancer research.
The core principle comes from language processing. In large language models, in-context learning has become standard, allowing a system to perform a new task from just a few examples without retraining. PRET applies this concept to histopathological images. To recognize a new cancer type, the system needs only one to eight annotated tissue samples as reference. The system then adjusts its outputs without changing its underlying parameters.
Traditional medical AI systems are trained on large datasets from a specific institution or patient population. Training takes weeks to months and must be repeated whenever the scanner changes, the patient population shifts, or a hospital switches. PRET bypasses this step. The system was pre-trained on a broad spectrum of histopathological images and can handle new tasks through in-context learning without adjusting its base state.
Why the retraining problem in pathology is real
Tissue samples look different across institutions. Different staining methods, tissue preparation protocols, scanner technology, and patient populations mean a model trained at Hospital A often performs worse at Hospital B or needs recalibration. That requires time, expertise, and resources many hospitals worldwide cannot provide.
PRET solves this with its architecture. Researchers validated the system across 23 international benchmark datasets from medical institutions in mainland China, the United States, and the Netherlands, reaching a consistent result: the system works without retraining across different institutional contexts.
23 datasets, 20 tasks, 18 cancer types
Across the 23 datasets, researchers tested PRET on 20 diagnostic tasks: cancer screening, tumor classification, tumor segmentation, and detection of lymph node metastases. In 20 of these tasks, PRET outperformed previously published systems. In 15 tasks, the system achieved an AUC value above 97 percent.
For colorectal cancer screening, PRET achieved an AUC of 100 percent according to HKUST data. For lymph node detection, the system reached 98.71 percent. Eleven pathologists performing the same task averaged 81 percent. This figure should be read carefully: clinical pathologists have access to far more context than individual tissue samples when making diagnoses. The result shows, however, that PRET performs at clinically relevant levels for this specific step of image analysis.
In comparison: Other AI systems in digital pathology
PRET is not the first AI system in digital pathology. Paige AI, a New York-based company, received FDA approval in September 2021 for the first AI system for prostate cancer diagnosis. In clinical trials, Paige Prostate increased pathologist sensitivity for tumor detection from 88.7 to 96.6 percent and reduced false negatives by about 70 percent. PathAI and Ibex Medical Analytics offer comparable systems for specific cancer types.
The structural difference with PRET is clear. These systems are optimized for single cancer types and specific institutional contexts. Each new application or institution requires substantial retraining. PRET covers 18 cancer types at once and starts without that effort. The research team sees the system as the foundation for universally deployable pathology diagnostics, especially in regions where specialized facilities and experienced pathologists are scarce.
From lab to clinic: What PRET still needs
Between publication in Nature Cancer and routine use in hospitals typically lies two to five years. In the United States, the FDA oversees medical software approval; in Europe, the European Medicines Agency works with national authorities. This requires prospective clinical trials testing the system under real-world conditions with clearly defined patient populations. Paige AI was initially designated a Breakthrough Device in 2019 and received regular FDA approval in September 2021.
Second, PRET must integrate into existing clinical workflows, requiring interfaces with digital pathology scanners and hospital information systems. Third, broad adoption needs physicians who can interpret system outputs and question them when needed. AI systems in pathology are not autonomous diagnostic authorities. They direct attention to relevant tissue areas. Final diagnosis and clinical decision remain with the physician.