A research team has developed an artificial intelligence pathology system capable of accurately identifying multiple cancer types using only a minimal number of annotated samples, with no additional training required for each new diagnostic task. The system, developed at the Hong Kong University of Science and Technology in collaboration with medical institutions in mainland China and the United States, represents a significant departure from how AI diagnostic tools have traditionally been built and deployed.
Conventional models used in pathology typically require tens of thousands of images to train for each specific cancer type, resulting in lengthy development timelines and substantial computational costs. That requirement has limited practical use in clinical settings, particularly in regions with fewer resources. The new system sidesteps that barrier by borrowing a concept from natural language processing called in-context learning, which allows it to adapt to a new cancer type or diagnostic task by referencing as few as one to eight annotated tumor slides at the time of use rather than undergoing a separate training process for each application.
Performance across 18 cancer types
The system was validated against 23 international benchmark datasets drawn from medical institutions across mainland China, the United States, and the Netherlands, covering 18 cancer types and a range of diagnostic tasks including cancer screening, tumor classification, and tumor boundary identification. It outperformed existing methods in 20 of the tasks evaluated, and its diagnostic accuracy exceeded 97 percent in 15 of those tasks.
In colorectal cancer screening, the system achieved a perfect accuracy score. In esophageal cancer tumor segmentation it reached 99.54 percent. The most striking result came in lymph node metastasis detection, one of the most technically demanding tasks in pathology, where the system attained an accuracy of approximately 98.71 percent using only eight slide samples. A panel of 11 pathologists performing the same task averaged roughly 81 percent accuracy, placing the AI system meaningfully above the human benchmark on a test that directly affects treatment decisions for cancer patients worldwide.
A tool for global health equity
Nearly 20 million new cancer diagnoses are made globally each year, and pathological examination plays a central role in determining how those cases are treated. At the same time, a severe global shortage of qualified pathologists has placed growing pressure on health systems to find ways to expand diagnostic capacity without proportionally expanding the workforce. Artificial intelligence has long been proposed as part of the solution, but the training requirements of conventional systems have made deployment difficult in exactly the settings where the need is greatest.
The plug-and-play design of the new system addresses that constraint directly. Because it does not require task-specific fine-tuning, it can be applied across different tumor types and clinical environments without the infrastructure investment that existing tools demand. Researchers described its core value as breaking down the traditional barriers of large datasets and repetitive training cycles, making advanced diagnostic capability accessible at lower cost and with greater flexibility than what has previously been possible.
What comes next
The research findings were published in the journal Nature Cancer. The team plans to extend the system’s capabilities to additional clinical applications including genetic mutation prediction and patient prognosis assessment, expanding the range of decisions it can support beyond initial diagnosis.
The broader ambition behind the project is to make precision diagnostic services available regardless of geography or resource level, addressing one of the most persistent structural inequities in global cancer care. Whether that ambition translates into widespread clinical adoption will depend on regulatory pathways, institutional integration, and further validation across diverse patient populations, but the performance results reported so far establish a compelling foundation.
