HKUMed Develops AI Model Achieving Over 90% Accuracy in Thyroid Cancer Diagnosis
Image Credit: National Cancer Institute | Splash
Researchers at the University of Hong Kong’s Li Ka Shing Faculty of Medicine (HKUMed), in collaboration with the InnoHK Laboratory of Data Discovery for Health and the London School of Hygiene & Tropical Medicine, have developed an artificial intelligence model to support thyroid cancer diagnosis. The findings were published in the journal npj Digital Medicine.
AI Application in Thyroid Cancer Classification
The research team reported that the AI tool can classify thyroid cancer stage and recurrence risk with an accuracy rate exceeding 90%. The model was trained using 50 pathology reports from The Cancer Genome Atlas (TCGA) and further validated with 289 additional TCGA reports and 35 simulated clinical cases prepared by endocrine surgeons.
The system integrates four open-source large language models—Mistral, Llama, Gemma, and Qwen—to process pathology reports and simulated clinical notes. Using a majority-vote approach, the model achieved 92.9% to 98.1% accuracy for American Joint Committee on Cancer (AJCC) staging and 88.5% to 100% accuracy for American Thyroid Association (ATA) risk classification, according to data in the published study.
Efficiency and Data Privacy
According to the HKUMed team, the AI model can reduce the time required for doctors to review and prepare medical reports by approximately half. The model is designed to operate offline, which the researchers state can help protect patient data privacy and allow integration into hospital systems.
Clinical Context and Implementation
Thyroid cancer is the fifth most common cancer among women in Hong Kong, according to the Hong Kong Cancer Registry. Accurate staging and risk assessment are important for patient management, but manual review of records can be time-consuming and subject to variation between clinicians. The HKUMed team stated that the new AI model is intended to support standardisation and improve efficiency in clinical workflows.
Limitations and Next Steps
The study authors note that the model was trained predominantly on early-stage cancer data and may have limited applicability in advanced cases. The tool relies on clear and complete medical documentation for accurate analysis. The model is based on the 8th edition AJCC and previous ATA guidelines, and future updates may be required to align with current standards. The TCGA dataset used for training does not contain surgical or imaging data.
HKUMed researchers indicated that further validation with more diverse, real-world datasets is planned. The model is currently available as open-source software, enabling external evaluation and development.
TheDayAfterAI News
We are a leading AI-focused digital news platform, combining AI-generated reporting with human editorial oversight. By aggregating and synthesizing the latest developments in AI — spanning innovation, technology, ethics, policy and business — we deliver timely, accurate and thought-provoking content.