HKUST Researchers Unveil Plug-and-Play AI System Outperforming Pathologists in Multi-Cancer Detection Tasks

New plug-and-play AI from HKUST detects 18 cancer types with 97% accuracy, surpassing human pathologists in metastasis detection using minimal samples.

By: AXL Media

Published: Apr 22, 2026, 4:11 AM EDT

Source: Information for this report was sourced from Hong Kong University of Science and Technology

A Paradigm Shift in Diagnostic Artificial Intelligence

The Hong Kong University of Science and Technology, in collaboration with Harvard Medical School and Guangdong Provincial People’s Hospital, has unveiled a novel pathology analysis system named PRET. This technology, which stands for Pan-cancer Recognition without Example Training, represents a fundamental departure from traditional medical AI models that typically require tens of thousands of annotated images for each specific cancer type. By utilizing a "plug-and-play" architecture, the system can instantly adapt to new diagnostic tasks during the inference stage, effectively bypassing the lengthy and expensive development cycles that have historically hampered the adoption of intelligent pathology.

Technical Foundations Borrowed from Linguistic Models

At the heart of PRET is the integration of "in-context learning," a concept originally developed for natural language processing, now applied to the complex field of pathological image analysis. This innovation allows the model to reference as few as one to eight annotated tumor slides to perform cancer screening, tumor subtyping, and segmentation without requiring task-specific fine-tuning. According to Professor Li Xiaomeng, the lead researcher and Associate Director of the Center for Medical Imaging and Analysis at HKUST, this approach breaks down the traditional barriers of massive data dependency and repetitive training, enabling AI tools to be applied in clinical settings with unprecedented flexibility.

Global Validation Across Diverse Medical Institutions

The research team conducted a rigorous validation of the system using 23 international benchmark datasets sourced from medical institutions in the United States, the Netherlands, and mainland China. Covering 18 different cancer types, the PRET system demonstrated superior performance in 20 distinct clinical tasks. Statistical analysis showed that the model achieved an Area Under the Curve, a primary metric for diagnostic accuracy, exceeding 97% in 15 of these benchmarks. These findings, recently published in the journal Nature Cancer, suggest that the tool maintains high reliability across varying populations and regional medical infrastructures.