23-04-2025
University of Hong Kong medical school develops world's first AI model for thyroid cancer diagnosis
The University of Hong Kong's medical school (HKUMed) has developed the world's first AI model for diagnosing thyroid cancer, showing over 90 per cent accuracy and slashing clinicians' pre-consultation time by 50 per cent.
The new AI model, trained to analyse clinical documents, can classify the stage and risk category of thyroid cancer, HKUMed announced on Wednesday.
The medical school said its model is more efficient than the traditional manual integration of clinical information conducted through the systems of the American Joint Committee on Cancer (AJCC) and the American Thyroid Association (ATA).
Researchers trained the AI model with pathology reports of 50 thyroid cancer patients from The Cancer Genome Atlas Programme (TCGA) using four offline open-source large language models, including Google's Gemma and Meta's Llama.
The team then checked the results against pathology reports from 289 TCGA patients, as well as 35 pseudo-cases created by endocrine surgeons.
The accuracy exceeded 90 per cent in classifying cancer stages and risk categories, HKUMed said.
The AI assistant provides high accuracy in extracting and analysing information from complicated pathology reports, operation records, and clinical notes, said Matrix Fung, chief of endocrine surgery at HKUMed and one of the project's two leading researchers.
The model can also be 'readily integrated' into the public and private healthcare sectors, as well as local and overseas research institutes, Fung said.
'[O]ur AI model also dramatically reduces doctors' preparation time by almost half compared to human interpretation,' he said, adding that 'doctors will have more time to counsel with their patients.'
Professor Joseph Wu, another HKUMed academic leading the research, pointed out the model's offline capability as a major advantage, allowing doctors to use it without having to share or upload patients' information online, thus protecting patient privacy.
HKUMed said the next step would be reviewing the performance of the AI assistant with a large amount of real-world patient data.
The model can be deployed in real clinical settings and hospitals once the results are validated, it said.
