文献速递甲状腺与浅表头颈US

基于ChatGPT-5的大语言模型与FDA批准的AI-CAD系统用于甲状腺结节超声评估的对比

ChatGPT-5-Based large language model analysis versus an FDA-approved AI-CAD system for thyroid nodule ultrasound evaluation.
PubMed 文献 · Eur J Radiol · 2025-12-31

前瞻性纳入141例患者186个甲状腺结节,比较ChatGPT-5与S-Detect基于超声图像良恶性分类与病理的一致性。ChatGPT-5敏感度50.8%、特异度75.8%、准确率59.1%,S-Detect分别为91.9%、62.9%、82.3%;S-Detect的AUC 77.4%显著高于ChatGPT-5的63.3%(P < 0.001)。结论认为ChatGPT-5用于甲状腺结节分类可行,但诊断性能低于S-Detect,尚不适用于医学影像应用。

要点

  • S-Detect与病理一致性中等(Kappa=0.579),ChatGPT-5一致性一般(Kappa=0.224)。
  • S-Detect敏感度91.9%、准确率82.3%,高于ChatGPT-5的50.8%和59.1%。
  • S-Detect的AUC 77.4%显著高于ChatGPT-5的63.3%(P < 0.001)。

核心结论

  • ChatGPT-5用于甲状腺结节分类具有可行性,但诊断性能低于经许可的预训练S-Detect系统,尚不适用于医学影像应用。

研究设计与样本

  • 前瞻性研究纳入141例患者186个结节,均行术前超声及后续手术;S-Detect分析每个结节最大横切灰阶超声图像并行自动勾画二分类,ChatGPT-5分析裁剪后的横切与纵切结节超声图像并使用标准化诊断提示词行二分类;以组织病理学为参考标准,采用Kappa统计评估一致性并计算敏感度、特异度、准确率及AUC。

诊断标准与关键数值

  • ChatGPT-5敏感度50.8%、特异度75.8%、准确率59.1%、AUC 63.3%、与病理Kappa=0.224;S-Detect敏感度91.9%、特异度62.9%、准确率82.3%、AUC 77.4%、Kappa=0.579;两系统区分良恶性结节均有统计学意义(P < 0.05),S-Detect的AUC显著大于ChatGPT-5(P < 0.001)。

临床应用与阅片要点

  • S-Detect的AUC 77.4%显著高于ChatGPT-5的63.3%(P < 0.001),ChatGPT-5尚不适用于医学影像应用。

局限性

  • 原文未报告局限性。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

英文原文(摘要)· 点击展开

PURPOSE: Recent advances in multimodal large language models (LLMs) have demonstrated promising potential for medical image analysis, yet their diagnostic capability in thyroid ultrasound remains unverified. This study explored the feasibility of ChatGPT-5, the latest multimodal LLM, for thyroid nodule classification and contextualized its diagnostic performance against S-Detect, an FDA-approved commercial computer-aided diagnosis system.

METHODS: In this prospective study, 141 patients with 186 nodules who underwent preoperative ultrasound and subsequent surgery were enrolled. For S-Detect, the largest transverse grayscale ultrasound image of each nodule was analyzed with automated contouring for binary classification. For ChatGPT-5, cropped transverse and longitudinal nodule ultrasound images were analyzed using a standardized diagnostic prompt for binary classification. Agreement with histopathology was assessed using Kappa statistics; sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve (AUC) were calculated.

RESULTS: Both systems showed statistically significant ability to distinguish benign from malignant nodules (P < 0.05). Agreement with histopathology was fair for ChatGPT-5 (Kappa = 0.224) and moderate for S-Detect (Kappa = 0.579). ChatGPT-5 demonstrated sensitivity 50.8 %, specificity 75.8 %, and accuracy 59.1 %, whereas S-Detect achieved higher sensitivity (91.9 %) and accuracy (82.3 %) but lower specificity (62.9 %). The AUC for S-Detect (77.4 %) was significantly greater than that for ChatGPT-5 (63.3 %, P < 0.001).

CONCLUSIONS: ChatGPT-5 demonstrated feasibility for thyroid nodule classification but showed lower diagnostic performance than the licensed, pre-trained S-Detect system and is not yet adequate for medical imaging applications.

来源:Eur J Radiol,查看原文。版权归原作者所有。

引用

  1. Chen Z, Ye M, Liang J, et al. ChatGPT-5-Based large language model analysis versus an FDA-approved AI-CAD system for thyroid nodule ultrasound evaluation. Eur J Radiol. 2025. PMID: 41494445. doi:10.1016/j.ejrad.2025.112639

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。