'ThyGPT' slashes rates of thyroid nodule biopsies
Experts have developed a specialized generative pre-trained transformer model they believe could significantly improve the thyroid nodule risk assessment process.
A new paper in Nature: NPJ Digital Medicine details the efforts of a group of researchers with the Chinese Academy of Sciences in Hangzhou, China, to develop and train a generative pre-trained transformer for thyroid nodules. The team says the final product—dubbed ThyGPT—is capable of significantly reducing biopsy rates without decreasing reader accuracy.
“Although using artificial intelligence to analyze ultrasound images is a promising approach to assessing thyroid nodule risks, traditional AI models lack transparency and interpretability,” Dong Xu, with the Department of Diagnostic Ultrasound Imaging & Interventional Therapy at Zhejiang Cancer Hospital, and colleagues explained. “We developed a multimodal generative pre-trained transformer for thyroid nodules, aiming to provide a transparent and interpretable AI copilot model for thyroid nodule risk assessment and management.”
Integrated with large language models and computer vision techniques, the model was trained using ultrasound data on thyroid nodules from nearly 60,000 patients across 9 hospitals. The ThyGPT’s training integrated imaging, error-free text reports, clinical histories, pathological data and diagnostic guidelines to produce the most thorough risk assessments possible.
The model underwent two testing assessments—one to determine its accuracy and efficacy for reducing unnecessary biopsies and a second to test its ability to spot errors in radiology reports.
For the first test, ThyGPT was used by three junior radiologists and three senior radiologists (more than 10 years of experience) who were assessing ultrasound images. During this assessment, the team found that when radiologists engaged in communication with ThyGPT, their accuracy increased significantly. With the improvement of readers’ sensitivity and specificity, the group determined that use of the model would have resulted in a 40% reduction in unnecessary nodule biopsies.
In the second assessment, radiologists were tasked with detecting and categorizing errors in radiology reports. This revealed ThyGPT to be more effective at spotting errors compared to human radiologists. However, with the help of ThyGPT, radiologists’ error detection rates also substantially improved, with junior readers seeing the biggest boost in performance. What’s more, ThyGPT took less than half a second to arrive at its conclusions, while radiologists averaged 49 seconds, representing significant time savings.
“The key advantage of ThyGPT is its ability to engage in natural language interaction, enabling radiologists to query the model’s rationale and obtain detailed explanations for its diagnoses,” the group noted. “This transparency addresses a major limitation of existing CAD models, which often function as ‘black boxes’ and ‘mute boxes.’”
The model’s transparency fosters trust and confidence in its diagnoses, which has been a big concern for radiologists who are hesitant to adopt the technology, the group indicated. They suggested that models that “show their work” are more likely to reach clinical settings and, thus, improve workflows for both providers and patients.
Learn more about ThyGPT here.
In addition to her background in journalism, Hannah also has patient-facing experience in clinical settings, having spent more than 12 years working as a registered rad tech. She began covering the medical imaging industry for Innovate Healthcare in 2021.