Scholarlink

With the continuous development of artificial intelligence (AI), the popularity and performance of Large Language Models (LLMs) have been improving. At present, there is a lack of experimental evaluation of LLMs for auxiliary clinical diagnosis. This study employs an analysis of thyroid nodule imaging evaluation to investigate the practical capabilities of LLMs in providing clinical decision support within radiology. To evaluate LLMs, including DeepSeek, Kimi, Doubao and ERNIE Bot, this study input 15 descriptive statements of thyroid nodules designed based on the Thyroid Imaging, Reporting and Data System launched by the American College of Radiology (ACR TI-RADS) and 241 ultrasound reports collected from a hospital, and asked LLMs to provide corresponding ACR TI-RADS grades. By comparing ACR TI-RADS with the output given by LLMs, the grading accuracy of LLMs’ output was tested. The study was repeated three times. All LLMs demonstrated perfect accuracy (100%) when evaluating standardized descriptive statements of thyroid nodules. However, a decline in diagnostic concordance was observed during the assessment of clinical ultrasound reports, with accuracy quantified as follows: DeepSeek 81.74 ± 2.77%, Kimi 61.55 ± 11.64%, Doubao 82.71 ± 3.09% and ERNIE Bot 83.26 ± 6.25%. The result shows that ERNIE Bot has the best overall performance. In general, LLMs have clinical potential in evaluating thyroid nodules.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

The Application Value of Large Language Models in ACR TI-RADS Grading Evaluation of Thyroid Nodules

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Lisheng Che,
  • Ranran Dai,
  • Penghui Wang,
  • Zhiheng Xu,
  • Wei Wang,
  • Wenxian Peng

摘要

With the continuous development of artificial intelligence (AI), the popularity and performance of Large Language Models (LLMs) have been improving. At present, there is a lack of experimental evaluation of LLMs for auxiliary clinical diagnosis. This study employs an analysis of thyroid nodule imaging evaluation to investigate the practical capabilities of LLMs in providing clinical decision support within radiology. To evaluate LLMs, including DeepSeek, Kimi, Doubao and ERNIE Bot, this study input 15 descriptive statements of thyroid nodules designed based on the Thyroid Imaging, Reporting and Data System launched by the American College of Radiology (ACR TI-RADS) and 241 ultrasound reports collected from a hospital, and asked LLMs to provide corresponding ACR TI-RADS grades. By comparing ACR TI-RADS with the output given by LLMs, the grading accuracy of LLMs’ output was tested. The study was repeated three times. All LLMs demonstrated perfect accuracy (100%) when evaluating standardized descriptive statements of thyroid nodules. However, a decline in diagnostic concordance was observed during the assessment of clinical ultrasound reports, with accuracy quantified as follows: DeepSeek 81.74 ± 2.77%, Kimi 61.55 ± 11.64%, Doubao 82.71 ± 3.09% and ERNIE Bot 83.26 ± 6.25%. The result shows that ERNIE Bot has the best overall performance. In general, LLMs have clinical potential in evaluating thyroid nodules.