Innovative AI Model Achieves Over 90% Accuracy in Thyroid Cancer Diagnosis

A groundbreaking interdisciplinary team of researchers has developed the world’s first artificial intelligence (AI) model capable of accurately classifying both the stage and risk level of thyroid cancer, surpassing 90% accuracy. This cutting-edge AI system utilizes four offline large language models—Mistral, Llama, Gemma, and Qwen—to analyze clinical documents efficiently for precise cancer staging and risk assessment.

The model was trained on open-access U.S. pathology reports from the Cancer Genome Atlas Program (TCGA), covering 50 thyroid cancer cases, and validated with reports from 289 patients, along with 35 pseudo-cases created by endocrine surgeons. By integrating the outputs of these models, the system achieves an impressive accuracy range: 88.5% to 100% for ATA risk classification and 92.9% to 98.1% for AJCC staging.

This innovation is expected to significantly reduce clinicians' pre-consultation preparation time by approximately half, enhancing workflow efficiency and patient care. Professor Joseph T Wu from HKUMed highlighted the model’s superior performance and privacy benefits, noting that its offline functionality allows deployment without sharing patient data.

Compared with online large language models like GPT-4o and DeepSeek, the AI demonstrated comparable performance in tests, underscoring its potential for real-world application. Dr. Matrix Fung Man-him emphasized its high accuracy, efficiency, and versatility, noting that it can be integrated seamlessly into various healthcare settings worldwide.

The AI system can analyze complex pathology reports, operation notes, and clinical documents to provide accurate staging and risk stratification — critical tools for prognosis and treatment planning. Future steps include validating the model’s performance with extensive real-world data, paving the way for deployment in clinical settings to improve operational efficiency and patient outcomes.