Multimodal GPT model for assisting thyroid nodule diagnosis and management

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Traditional artificial intelligence (AI) models for analyzing thyroid ultrasound images lack transparency and interpretability.
• Accurate risk assessment of thyroid nodules is crucial for appropriate patient management.

Purpose of the Study:

• To develop and evaluate ThyGPT, a multimodal generative pre-trained transformer for thyroid nodules.
• To provide a transparent and interpretable AI copilot for thyroid nodule risk assessment and management.

Main Methods:

• Retrospective collection of ultrasound data from 59,406 patients across nine hospitals.
• Development and testing of the ThyGPT model, an AI-generated content-enhanced computer-aided diagnosis (AIGC-CAD) model.

Main Results:

• ThyGPT assisted in reducing biopsy rates by over 40% without increasing missed diagnoses.
• The AI copilot detected errors in ultrasound reports 1,610 times faster than human review.
• Radiologists' diagnostic performance improved, with the area under the curve increasing from 0.805 to 0.908 (p < 0.001).

Conclusions:

• ThyGPT offers a transparent and interpretable AI solution for thyroid nodule risk assessment.
• The AIGC-CAD model has the potential to significantly enhance the capabilities of radiologists in managing thyroid nodules.