Scholarlink

Introduction <p>The accuracy of radiology reports is essential to guide clinical decisions and ensure high-quality patient care. However, all stages of the interpretive process remain prone to various errors, including transcription errors. This study aimed to assess the performance of ChatGPT in the linguistic and clinical analysis of radiology reports, comparing its results with those of radiology residents.</p> Materials and methods <p>A retrospective experimental study was conducted between November 2024 and April 2025. One hundred radiology reports were randomly divided into two groups: one without errors and the other with 100 deliberately introduced errors. The reports included 50 standard radiographs and 50 computed tomography scans. Four radiology residents independently reviewed all 100 reports and GPT-4 was subjected to the same documents.</p> Results <p>GPT-4 detected 71% of errors (95% CI 61–79), compared with 52% (95% CI 42–62) for residents, with a statistically significant difference (<i>p</i> = 0.008). GPT-4 significantly outperformed residents in the identification of spelling, laterality, and other types of errors (<i>p</i> &lt; 0.05). Furthermore, GPT-4 showed higher sensitivity (87% vs. 76%), specificity (93% vs. 75%), and accuracy (94% vs. 74%), with a substantially shorter average review time per report (28&#xa0;s vs. 111&#xa0;s).</p> Conclusion <p>GPT-4 outperforms radiology residents in detecting errors in radiology reports and could serve as a valuable support tool to improve radiological care and reduce workload.</p>

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

Improving the accuracy of radiology reports with ChatGPT: evaluation of its performance in error detection and correction

  • 期刊论文

  • 发表日期：

    2026年1月8日

查看全文
  • Pihou Gbande,
  • Jean Modeste Ngoune,
  • Mazamaesso Tchaou,
  • Lantam Sonhaye,
  • Lama Kegdigoma Agoda-Koussema,
  • Komlanvi Adjenou

摘要

Introduction

The accuracy of radiology reports is essential to guide clinical decisions and ensure high-quality patient care. However, all stages of the interpretive process remain prone to various errors, including transcription errors. This study aimed to assess the performance of ChatGPT in the linguistic and clinical analysis of radiology reports, comparing its results with those of radiology residents.

Materials and methods

A retrospective experimental study was conducted between November 2024 and April 2025. One hundred radiology reports were randomly divided into two groups: one without errors and the other with 100 deliberately introduced errors. The reports included 50 standard radiographs and 50 computed tomography scans. Four radiology residents independently reviewed all 100 reports and GPT-4 was subjected to the same documents.

Results

GPT-4 detected 71% of errors (95% CI 61–79), compared with 52% (95% CI 42–62) for residents, with a statistically significant difference (p = 0.008). GPT-4 significantly outperformed residents in the identification of spelling, laterality, and other types of errors (p < 0.05). Furthermore, GPT-4 showed higher sensitivity (87% vs. 76%), specificity (93% vs. 75%), and accuracy (94% vs. 74%), with a substantially shorter average review time per report (28 s vs. 111 s).

Conclusion

GPT-4 outperforms radiology residents in detecting errors in radiology reports and could serve as a valuable support tool to improve radiological care and reduce workload.