文献速递胸部CTDR

视觉解析与诊断推理的解耦:基于胸部影像测验病例分析视觉语言模型(GPT-4o与GPT-5)

Decoupling Visual Parsing and Diagnostic Reasoning for Vision-Language Models (GPT-4o and GPT-5): Analysis Using Thoracic Imaging Quiz Cases.
PubMed 文献 · AJR Am J Roentgenol · 2025-12-10

本回顾性研究纳入128例公开胸部影像测验病例,评估GPT-4o与GPT-5在输入图像与输入放射科医师描述两种条件下的诊断表现;两模型视觉解析能力有限,但在提供放射科医师描述后诊断准确性提高,提示视觉图像解析而非诊断推理是限制模型表现的主要环节。

要点

  • GPT-4o与GPT-5在仅输入图像时top-5准确率分别为15.9%与24.7%,低于测验参与者汇总的45.8%。
  • 提供放射科医师图像描述后,两模型top-5准确率分别升至40.1%与59.1%。
  • 两模型图像摘要评分中位数均为2分,评分与top-5匹配独立相关(GPT-4o OR=5.95;GPT-5 OR=2.77;p<.001)。

核心结论

  • 两种视觉语言模型视觉识别胸部影像征象的能力有限,但在获得放射科医师生成的描述后生成正确诊断的表现更好;视觉图像解析而非诊断推理是限制模型表现的主要环节。

研究设计与样本

  • 回顾性研究,纳入韩国胸部影像学会测验平台128例公开胸部影像病例(2025年6月15日访问),GPT-4o与GPT-5分别在接受患者元数据加图像、以及患者元数据加放射科医师图像描述两种条件下处理病例,以模型五个排序鉴别诊断是否包含正确诊断作为top-5准确率。

诊断标准与关键数值

  • GPT-4o与GPT-5在输入元数据加图像时top-5准确率分别为15.9%与24.7%,输入元数据加描述时分别为40.1%与59.1%;测验参与者汇总top-5准确率为45.8%;模型图像摘要评分中位数均为2分(4分制),摘要评分与top-5匹配独立相关(GPT-4o OR=5.95;GPT-5 OR=2.77;p<.001)。

临床应用与阅片要点

  • 作者指出结果强调放射科医师在胸部影像判读中的专业价值,并识别视觉图像解析而非诊断推理是限制视觉语言模型表现的主要环节。

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

BACKGROUND. Vision-language models (VLMs) have potential to identify findings on radiologic imaging (i.e., visual parsing) and translate findings into diagnoses (i.e., diagnostic reasoning). Current VLMs have shown insufficient performance to support clinical integration. OBJECTIVE. The purpose of our study was to evaluate the separate contributions of visual parsing and diagnostic reasoning toward GPT-based VLMs' performance in generating correct diagnoses for thoracic imaging. METHODS. This retrospective study included 128 publicly available thoracic imaging cases from the Korean Society of Thoracic Imaging quiz platform (accessed on June 15, 2025). Two VLMs (GPT-4o and GPT-5) processed cases, separately when patient metadata and images were inputted and when patient metadata and radiologist-generated image descriptions were inputted. The models provided five ranked differential diagnoses for each case; when metadata and images were inputted, the models first provided a summary of imaging findings. The proportion of cases for which the models' five differential diagnoses included the correct diagnosis was determined (i.e., top-5 accuracy). The performance of quiz participants, who interpreted cases using metadata and images, was extracted from the platform. The quality of the model-provided image summaries was scored on a 4-point scale (4 = best score). Logistic regression analyses assessed associations between model image summary scores and diagnostic performance. Diagnostic concordance was assessed between models' top-ranked diagnoses and quiz participants' top-10 differential diagnoses. RESULTS. Top-5 accuracy for GPT-4o and GPT-5 when metadata and images were inputted was 15.9% and 24.7% and when metadata and descriptions were inputted was 40.1% and 59.1%, respectively; quiz participants' pooled top-5 accuracy was 45.8%. Median image summary score was 2 for both models; these scores showed significant independent associations with a top-5 match (GPT-4o: OR = 5.95; GPT-5: OR = 2.77; p < .001). Concordance between models' top-ranked diagnosis and quiz participants' differential lists for GPT-4o and GPT-5 when metadata and images were inputted was 31.6% and 39.3% and when metadata and descriptions were inputted was 78.8% and 79.4%, respectively. CONCLUSION. Two VLMs showed limited ability to visually identify thoracic imaging findings but performed more favorably in generating accurate diagnoses when provided radiologist-generated descriptions. CLINICAL IMPACT. The results underscore the need for radiologist expertise in thoracic imaging interpretation and identify visual image parsing rather than diagnostic reasoning as the principal limitation constraining VLM performance.

来源:AJR Am J Roentgenol,查看原文。版权归原作者所有。

引用

  1. Han DH, Hwang EJ, Yoon SH, et al. Decoupling Visual Parsing and Diagnostic Reasoning for Vision-Language Models (GPT-4o and GPT-5): Analysis Using Thoracic Imaging Quiz Cases. AJR Am J Roentgenol. 2025. PMID: 41370655. doi:10.2214/AJR.25.34116

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。