Photo: Longhua Liao/Getty Images
Researchers from the Hong Kong University of Science and Technology have developed a new AI-powered pathology analysis system that identifies multiple cancer types and performs various tasks using a few slides without the need for additional training.
In partnership with Guangdong Provincial People's Hospital and Harvard Medical School, the team created PRET (Pan‑cancer Recognition without Example Training), which it said is the first system to introduce the concept of "in-context learning" from natural language processing into pathological image analysis.
HOW IT WORKS
In-context learning allows the model to instantly adapt to new cancer types and perform diagnostic tasks during inference by referencing a handful of annotated slides.
"Our model leverages local image patches as contextual cues," research lead and HKUST assistant professor LI Xiaomeng told Mobihealth News.
"By treating these local patches as in-context examples, the model is able to effectively exploit fine-grained local information. This process is concretely implemented through matrix operations within the model, which enable it to perform adaptive inference without the need for parameter updates or additional training," he explained.
He also mentioned that PRET is model-agnostic. "[I]t does not tie itself to a specific base model and can thus be flexibly applied to extend any existing pathology foundation model."
FINDINGS
The research team validated PRET on 23 international benchmark datasets from China, the United States, and the Netherlands, covering 18 cancer types and various diagnostic tasks, such as cancer screening, tumour subtyping, and tumour segmentation.
Findings published in Nature Cancer showed that this system outperformed existing methods in 20 tasks, including colorectal cancer screening (100% area under the curve) and oesophageal squamous cell carcinoma tumour segmentation (99.54% AUC). Researchers also noted that the system achieved an AUC of 98.71% in detecting lymph node metastasis using only eight slides, surpassing the average performance of 11 pathologists at 81%.
Moreover, it demonstrated stable and robust generalisability across different populations and regions with varying levels of medical resources, the team said.
When asked about data bias and leakage, Prof Xiaomeng said the team took rigorous measures. "The vast majority of our validation datasets are sourced from newly scanned data provided by our collaborating hospitals. These specific datasets were not publicly available prior to our study and were consequently unseen by any models during their pre-training phases. Therefore, there is no risk of data leakage in our validation setup."
Researchers also noted the present limitations of their model, particularly its ability to distinguish between tumours with morphologically similar appearances.
WHY IT MATTERS
The HKUST researchers developed PRET to address limitations of existing AI models, which typically require volumes of training data and hefty computational and manpower costs. Conventional models also lack sufficient generalisability, needing extensive fine-tuning in live clinical settings.
PRET, designed as a plug-and-play diagnostic tool, does away with additional training, indicating potential for application in resource-constrained settings, the team said.
The model also exhibits better generalisation under low-data regimes, Prof Xiaomeng added, as it is less prone to overfitting. "It also excels in leveraging local image context."
The team plans to improve PRET's diagnostic performance and expand its applications to more clinical tasks, such as genetic mutation prediction and patient prognosis assessment.
The AI pathology analysis model, which has been open-sourced, according to Prof Xiaomeng, has yet to be piloted in clinical settings and deployed in hospital workflows.
THE LARGER TREND
HKUST has also developed a large language model-based pathology assistant tool, called mSTAR, which has several clinical diagnostic and prognostic applications, such as cancer subtyping and staging, metastasis detection, molecular prediction, survival analysis, and report generation.
Another AI-powered system developed at HKUST, SmartPath, automates pathology workflow and is trained on two large AI models and more than 500,000 whole-slide images across 34 cancer types.
Outside Hong Kong, SingHealth in Singapore is preparing to integrate more AI tools across its pathology services as it transitions to AI-enabled digital pathology.
