Khaberni - Researchers have developed a new system for pathology analysis using artificial intelligence, capable of identifying multiple types of cancer using a very limited number of samples, without the need for additional training.
This significant step may contribute to accelerating cancer diagnosis and improving the efficiency of health care globally.
This achievement comes at a time when about 20 million new cancer cases are diagnosed annually worldwide, while histological examination plays a fundamental role in clinical diagnosis and determining treatment plans. With a severe global shortage of pathologists, there is an increasing need for innovative technical solutions to enhance the efficiency of histological analysis and reduce pressure on medical staff.
Despite the significant potential offered by artificial intelligence in automating diagnosis, traditional models still face major challenges, typically requiring training on tens of thousands of images and data for each type of cancer or diagnostic task, which prolongs development time and increases human and computational costs. Moreover, these models often need fine-tuning when used with different types of tumors, limiting their scalability, especially in regions with limited resources.
To address these challenges, the research team led by the Hong Kong University of Science and Technology developed a new system named PRET, an acronym for "Pan-cancer Recognition Without Pre-training," in collaboration with the People's Hospital in Guangdong Province and Harvard Medical School.
PRET is the first system of its kind that applies the concept of "contextual learning" inspired by natural language processing in pathology image analysis. It can directly adapt to new types of cancer and perform multiple diagnostic tasks such as cancer screening, sub-type tumor classification, and tumor segmentation relying on just one to eight tissue slices previously identified and diagnosed by doctors, without the need for new training for each case.
This approach makes the system easier and more flexible compared to traditional models, as it eliminates the need for continuous adjustment for each diagnostic task, making it more suitable for practical use in hospital and real clinical environments.
The research team tested the system using 23 international reference datasets from medical institutions in China, the United States, and the Netherlands, including 18 different types of cancer and a variety of diagnostic tasks. The results showed PRET outperforming current methods in 20 tasks, with diagnostic accuracy exceeding 97% in 15 tasks.
Among the notable results, the system achieved 100% accuracy in colorectal cancer screening, and 99.54% in segmenting squamous cell carcinoma tumors in the esophagus. It also recorded an accuracy of about 98.71% in detecting lymph node metastases, one of the most complicated tasks, using only eight samples, thus surpassing the average performance of 11 pathology specialists, which was around 81%.
Professor Li Xiaoming, the leader of the research team, said: "The fundamental value of the PRET system lies in breaking the traditional barriers of massive data and repeated training, allowing the application of AI-supported pathology systems in real clinical environments at lower cost and with greater flexibility."
He added that this system not only reduces the workload on pathologists but can also improve access to accurate cancer diagnosis in regions deprived of medical services, thereby contributing to enhancing global health equity.
The team plans to further develop the system in the future, and expand its use to include additional clinical tasks such as predicting genetic mutations and assessing patients' recovery chances, opening up prospects for the future of disease diagnosis supported by artificial intelligence.
