文献速递甲状腺与浅表US

双语语音识别结合大语言模型辅助生成TI-RADS报告的两阶段流程

Streamlined TI-RADS Reporting From Bilingual Speech Recognition to LLM-Assisted Conclusion Generation Using a Two-Stage Workflow.
PubMed 文献 · Acad Radiol · 2026-09-01

本研究利用甲状腺影像数字数据库149例合格图像,评估基于大语言模型的双语语音识别与结论生成两阶段流程。结果显示,大语言模型语音识别在词错率、校正时间及医师偏好上均优于商用模型,详细提示策略显著提高TI-RADS评分准确性。作者强调仍需确定性后处理及前瞻性临床验证。

要点

  • 大语言模型语音识别词错率显著低于商用模型,可减少报告校正时间。
  • 采用详细提示策略可提高ACR TI-RADS与K-TIRADS评分准确性,但对阅片结论不能完全自动依赖。
  • 该流程为可行性研究,尚未经前瞻性临床验证,日常报告应结合确定性后处理与人工复核。

核心结论

  • 该两阶段流程中大语言模型语音识别效率优于商用模型,详细提示策略显著提高TI-RADS评分准确性。
  • 确定性后处理仍不可或缺,部署前必须完成前瞻性临床验证。

研究设计与样本

  • 研究类型为可行性研究,共纳入Digital Database of Thyroid Images数据集中图像质量合格的149例。
  • 由1名放射科医师生成金标准报告并口述报告,实施两阶段人工智能工作流程。
  • 第一阶段比较商用与大语言模型语音识别模型生成草稿报告,由2名放射科医师评估校正时间、偏好及词错率。
  • 第二阶段以GPT-4o从金标准报告生成TI-RADS评分,采用最小提示、详细提示及详细提示联合检索增强生成三种策略。

诊断标准与关键数值

  • 大语言模型语音识别总体词错率显著低于商用模型,分别为1.0 [IQR 0-2.0]与3.0 [IQR 2.0-3.0],P < 0.001。
  • 放射科医师使用大语言模型语音识别的校正时间显著更短,分别为6.38 [IQR 4.58-10.45]秒与11.39 [IQR 9.56-14.06]秒,P < 0.001。
  • 两位放射科医师均显著偏好大语言模型语音识别模型,偏好比例分别为70.5%与73.7%,P < 0.001。
  • 详细提示策略使ACR TI-RADS评分点准确性从最小提示的50.3%提高至78.5%,P < 0.001。
  • K-TIRADS分类准确性从58.4%提高至79.9%,P < 0.001。

临床应用与阅片要点

  • 日常报告可考虑用大语言模型语音识别生成草稿,但必须由放射科医师校正后再使用。
  • 采用包含完整指南的详细提示策略比简单提示更能保证甲状腺结节风险分层评分准确性。
  • TI-RADS自动生成结论应经过确定性后处理,避免大语言模型直接输出最终诊断结果。

局限性

  • 本研究为可行性研究,尚无前瞻性临床验证。
  • 样本来源于单一年份的公共图像数据库,样本量为149例,图像质量合格者才纳入。

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

RATIONALE AND OBJECTIVES: To evaluate a two-stage workflow for large language model (LLM)-assisted Thyroid Imaging Reporting and Data System (TI-RADS) reporting using LLM-based bilingual automated speech recognition (ASR) and conclusion generation.

MATERIALS AND METHODS: From the Digital Database of Thyroid Images dataset, 149 cases with adequate image quality were included. A radiologist generated gold standard reports and dictated the reports, implementing a two-stage artificial intelligence workflow. In stage 1, commercial and LLM-based ASR models were compared for generating draft reports. Two radiologists evaluated transcription performance using correction time, preference, and word error rate during manual correction. In stage 2, an LLM (GPT-4o) generated TI-RADS scores from gold standard reports using three prompting strategies: minimal (basic-scoring instruction), detailed (comprehensive-guidelines), and detailed with retrieval-augmented generation (RAG with guideline-integration). The accuracy of American College of Radiology (ACR) TI-RADS points and categories, and Korean Thyroid Imaging Reporting and Data System (K-TIRADS) categories was assessed. Statistical analyses included Wilcoxon matched-pairs signed-rank tests and McNemar's test.

RESULTS: In stage 1, the LLM-based ASR model demonstrated a significantly lower overall word error rate compared to the commercial ASR model (1.0 [IQR 0-2.0] vs 3.0 [IQR 2.0-3.0], P < 0.001). A radiologist required significantly shorter correction time with the LLM-based ASR model (6.38 [IQR 4.58-10.45] vs 11.39 [IQR 9.56-14.06] seconds, P < 0.001), and both radiologists showed a strong preference for the LLM-based ASR model (70.5% and 73.7%, P < 0.001). In stage 2, detailed prompting strategies significantly improved TI-RADS scoring accuracy. ACR TI-RADS point accuracy increased from 50.3% with minimal prompting to 78.5% with detailed prompting (P < 0.001). Korean TI-RADS category accuracy similarly improved from 58.4% to 79.9% (P < 0.001).

CONCLUSION: In this feasibility study of a two-stage workflow, the LLM-based ASR model demonstrated superior efficiency compared with the commercial model, and detailed prompting strategies significantly improved TI-RADS scoring accuracy; however, deterministic post-processing remained essential, and prospective clinical validation is required before deployment.

来源:Acad Radiol,查看原文。版权归原作者所有。

引用

  1. Han, Kim, Shin, et al. Streamlined TI-RADS Reporting From Bilingual Speech Recognition to LLM-Assisted Conclusion Generation Using a Two-Stage Workflow. Acad Radiol. 2026. PMID: 42680653. doi:10.1016/j.acra.2026.08.009

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。