文献速递技术与安全DR

深度伪造医学影像的兴起:放射科医生识别ChatGPT生成X线片的诊断准确性

The Rise of Deepfake Medical Imaging: Radiologists' Diagnostic Accuracy in Detecting ChatGPT-generated Radiographs.
PubMed 文献 · Radiology · 2026-01-01

这项2025年4月至8月开展的回顾性诊断准确性研究纳入来自6个国家的17名执业放射科医生,评估其与4种多模态大语言模型区分ChatGPT(GPT-4o)生成合成X线片与真实临床影像的能力。放射科医生对GPT-4o数据集的总体判别准确率为75%(95%CI:68,81),对RoentGen数据集为70%(95%CI:62,78),两者差异无统计学意义(P=0.07);GPT-4o(准确率85%)与GPT-5(准确率83%)的判别准确率高于Llama 4 Maverick(59%)和Gemini 2.5 Pro(56%)(均P<0.001)。作者认为合成X线片不易被放射科医生和大语言模型识别,培训医生与大语言模型识别合成图像对降低风险至关重要。

要点

  • 合成X线片常见双侧对称、颗粒或噪声模式均匀、软组织纹理细微不自然及骨表面过于光滑。
  • 放射科医生判别GPT-4o合成X线片的总体准确率为75%(95%CI:68,81),RoentGen数据集为70%(95%CI:62,78)。
  • 未设研究目的盲的17名放射科医生中41%(7/17)自发识别出数据集中存在人工智能生成X线片。

核心结论

  • 使用大语言模型生成的合成X线片(深度伪造)不易被放射科医生和大语言模型与真实X线片区分。
  • 作者认为培训医生与大语言模型识别合成图像对降低风险至关重要,并提供整理好的深度伪造数据集用于支持培训。

研究设计与样本

  • 回顾性诊断准确性研究,于2025年4月至8月开展,纳入来自6个国家的17名执业放射科医生,经验水平不一。
  • 第一阶段放射科医生对研究目的设盲,评估154张来自多个解剖部位的X线片(77张ChatGPT[GPT-4o;OpenAI]生成的合成图像与77张真实图像)的图像质量并给出诊断。
  • 第二阶段告知研究目的后,放射科医生判断随机呈现的X线片为GPT-4o生成还是真实图像,同一分类任务由GPT-4o、GPT-5(OpenAI)、Gemini 2.5 Pro(Google)和Llama 4 Maverick(Meta)四种大语言模型完成。
  • 第三阶段分析另外110张胸部X线片(55张RoentGen生成的合成图像与55张真实图像),评估读者与大语言模型区分合成与真实图像的表现,采用McNemar检验与t检验进行比较。

关键影像表现/征象

  • 合成X线片的常见特征包括双侧对称、颗粒或噪声模式均匀、软组织纹理细微不自然以及骨表面过于光滑。

诊断标准与关键数值

  • 17名放射科医生判别GPT-4o数据集合成图像的总体准确率为75%(95%CI:68,81),判别RoentGen数据集为70%(95%CI:62,78),两者差异无统计学意义(P=0.07)。
  • GPT-4o生成X线片被GPT-4o(准确率85%)和GPT-5(准确率83%)区分于真实图像的表现优于Llama 4 Maverick(准确率59%)和Gemini 2.5 Pro(准确率56%)(均P<0.001)。
  • 所测试的大语言模型在两个数据集中均未检出全部合成X线片。
  • 设盲于研究目的时,41%(7/17)的放射科医生自发识别出数据集中存在人工智能生成X线片。

临床应用与阅片要点

  • 阅片时可留意合成X线片常见特征:双侧对称、颗粒或噪声模式均匀、软组织纹理细微不自然及骨表面过于光滑。
  • 作者认为培训医生与大语言模型识别合成图像对降低风险至关重要,并提供了整理好的深度伪造数据集(https://noneedanick.github.io/DeepFakeXRay/)用于支持培训。

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 Large language models (LLMs) can generate realistic synthetic medical images (deepfakes), which raise concerns about potential misuse. Purpose To assess the ability of radiologists and multimodal LLMs to distinguish ChatGPT-generated synthetic radiographs from authentic clinical images. Materials and Methods This retrospective diagnostic accuracy study conducted between April and August 2025 included 17 practicing radiologists from six countries with varying experience levels. In phase 1, the radiologists, blinded to the purpose of the study, assessed image quality and provided diagnoses for 154 radiographs from multiple anatomic regions (77 synthetic images generated using ChatGPT [GPT-4o; OpenAI] and 77 authentic images). In phase 2, after being informed of the study's purpose, the radiologists determined whether randomly presented radiographs were GPT-4o-generated or authentic. The same classification task was performed by four LLMs: GPT-4o, GPT-5 (OpenAI), Gemini 2.5 Pro (Google), and Llama 4 Maverick (Meta). In phase 3, an additional set of 110 chest radiographs (55 synthetic images generated using RoentGen and 55 authentic images) was analyzed to evaluate the performance of readers and LLMs in distinguishing synthetic versus authentic images. The McNemar test and t test were used for comparisons. Results Forty-one percent (seven of 17) of purpose-blinded radiologists spontaneously identified artificial intelligence-generated radiographs as being present in the dataset. After being informed that some radiographs were synthetic, there was no evidence of a difference in overall accuracy among all 17 radiologists in distinguishing synthetic images in the GPT-4o dataset (75% [95% CI: 68, 81]) versus in the RoentGen dataset (70% [95% CI: 62, 78]; P = .07). No tested LLM detected all synthetic radiographs in either dataset; however, GPT-4o-generated radiographs were more accurately differentiated from authentic ones by GPT-4o (accuracy, 85%) and GPT-5 (accuracy, 83%) compared with Llama 4 Maverick (accuracy, 59%) and Gemini 2.5 Pro (accuracy, 56%) (all P < .001). Common features of synthetic radiographs included bilateral symmetry, uniform grain or noise patterns, subtly unnatural soft-tissue textures, and overly smooth bone surfaces. Conclusion Synthetic radiographs (deepfakes) generated using an LLM were not easily distinguishable from authentic radiographs by either radiologists or LLMs. Training physicians and LLMs to recognize synthetic images is essential to mitigate risks. To support training, a curated deepfake dataset is available: https://noneedanick.github.io/DeepFakeXRay/. © RSNA, 2026 Supplemental material is available for this article. See also the editorial by Bhayana and Krishna in this issue.

来源:Radiology,查看原文。版权归原作者所有。

引用

  1. Tordjman M, Yuce M, Ammar A, et al. The Rise of Deepfake Medical Imaging: Radiologists' Diagnostic Accuracy in Detecting ChatGPT-generated Radiographs. Radiology. 2026. PMID: 41874300. doi:10.1148/radiol.252094

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。