AI Breakthrough Outperforms Pathologists in Detecting Cancer Spread
AI Breakthrough Outperforms Pathologists in Detecting Cancer Spread
A research team from Hong Kong University of Science and Technology has unveiled a groundbreaking “plug-and-play” artificial intelligence system that can detect cancer and its spread more accurately than human pathologists—while requiring only a handful of sample data.
The new system, called PRET (Pan-cancer Recognition without Example Training), marks a major leap in AI-powered medical diagnostics. Unlike traditional AI models that need thousands of images and extensive training for each cancer type, PRET can analyze and recognize multiple cancers using as few as one to eight annotated samples.
Cancer remains one of the world’s most pressing health challenges, with nearly 20 million new cases diagnosed each year. Pathological analysis is critical for diagnosis and treatment decisions, but a global shortage of trained pathologists continues to strain healthcare systems. This is where PRET offers a promising solution.
Developed under the leadership of Li Xiaomeng in collaboration with Guangdong Provincial People's Hospital and Harvard Medical School, PRET introduces “in-context learning” to pathology—a concept adapted from natural language processing. This allows the system to quickly adapt to new diagnostic tasks without additional training.
In extensive testing across 23 global datasets covering 18 cancer types, PRET achieved remarkable results. It surpassed existing AI models in 20 diagnostic tasks and reached over 97% accuracy in most cases. In colorectal cancer screening, it achieved a perfect 100% accuracy rate.
Most notably, in detecting lymph node metastasis—a critical factor in cancer staging—PRET achieved an accuracy of about 98.7%, outperforming a group of 11 pathologists whose average accuracy was around 81%.
Researchers say the system could significantly reduce workload for medical professionals and improve access to accurate cancer diagnosis, especially in regions with limited resources. Its flexibility and low data requirement make it a strong candidate for real-world clinical use.
The study, published in the journal Nature Cancer, highlights the potential of AI to transform pathology and bring more equitable healthcare solutions worldwide.
Looking ahead, the team plans to expand PRET’s capabilities to include predicting genetic mutations and assessing patient prognosis—further pushing the boundaries of AI in medicine.
