World’s First AI Model for Thyroid Cancer Diagnosis Achieves Over 90% Accuracy

Thyroid cancer is one of the most common cancers worldwide, and its precise management typically relies on two primary systems: (1) the 8th edition of the American Joint Committee on Cancer (AJCC) or Tumor-Node-Metastasis (TNM) cancer staging system, which helps determine the stage of cancer, and (2) the American Thyroid Association (ATA) risk classification system, which is used to categorize cancer risk. These systems are vital for predicting patient survival and guiding treatment decisions. However, manually integrating the complex clinical data into these systems can be time-consuming and inefficient. Now, researchers have introduced the world’s first artificial intelligence (AI) model capable of classifying both the stage and risk category of thyroid cancer with remarkable accuracy, exceeding 90%. This innovative AI model, featured in the journal npj Digital Medicine, is set to significantly reduce pre-consultation preparation time for frontline clinicians by approximately 50%.

An interdisciplinary research team from the LKS Faculty of Medicine at the University of Hong Kong (HKUMed, Hong Kong) developed an AI assistant that employs large language models (LLMs) such as ChatGPT and DeepSeek. These models are designed to understand and process human language, helping to analyze clinical documents and improve the accuracy and efficiency of thyroid cancer staging and risk classification. The model utilizes four offline, open-source LLMs—Mistral (Mistral AI), Llama (Meta), Gemma (Google), and Qwen (Alibaba)—to interpret free-text clinical documents. The AI model was trained using an open-access data set from the United States that included pathology reports from 50 thyroid cancer patients from The Cancer Genome Atlas Program (TCGA), with validation done using pathology reports from 289 TCGA patients and 35 pseudo cases created by endocrine surgeons.

By combining the outputs from all four LLMs, the research team was able to enhance the overall performance of the AI model, achieving an accuracy rate ranging from 88.5% to 100% in ATA risk classification and 92.9% to 98.1% in AJCC cancer staging. In addition to its high accuracy in extracting and analyzing complex information from pathology reports, operation records, and clinical notes, the AI model also drastically reduces doctors’ preparation time by nearly half compared to manual interpretation. A significant advantage of this model is its ability to operate offline, allowing it to be deployed locally without the need to share or upload sensitive patient data, thereby ensuring maximum patient privacy. The AI model’s versatility means it can be easily integrated into a variety of healthcare settings, both public and private, as well as in international healthcare and research institutions.

“In line with government’s strong advocacy of AI adoption in healthcare, as exemplified by the recent launch of LLM-based medical report writing system in the Hospital Authority, our next step is to evaluate the performance of this AI assistant with a large amount of real-world patient data,” said Dr. Carlos Wong, Honorary Associate Professor in the Department of Family Medicine and Primary Care, School of Clinical Medicine, HKUMed. “Once validated, the AI model can be readily deployed in real clinical settings and hospitals to help clinicians improve operational and treatment efficiency.”

Related Links:
HKUMed