文献速递神经MRCT

GPT-4o脑肿瘤影像诊断可行性评估

Integrating large language models (LLMs) into neuro-oncologic radiology: a retrospective feasibility assessment of GPT-4o for brain tumor diagnosis.
PubMed 文献 · Neuroradiology · 2026-01-19

回顾性纳入239例术前病理证实的脑肿瘤病例,比较GPT-4o在三种输入方式下的诊断表现。仅图像输入的最终诊断准确率为32%,影像所见联合病史达76%,再加入图像后提升不明显。研究认为标准化文本输入可提升大语言模型诊断效能,但需进一步前瞻性验证。

要点

  • GPT-4o仅靠图像诊断脑肿瘤的最终准确率仅32%,不可单独用于诊断。
  • 结构化影像所见联合临床病史后,GPT-4o最终诊断准确率升至76%。
  • 加回图像对提高准确率无显著增益,提示文本质量是关键变量。

核心结论

  • 在术前脑肿瘤队列中,GPT-4o接受结构化影像所见和患者病史后的诊断准确率显著高于仅凭图像分析。
  • 标准化文本输入有提升GPT-4o诊断表现的潜在价值。
  • 该结果的临床普适性有限,需进一步前瞻性验证后方可考虑更广泛应用于临床。

研究设计与样本

  • 三种预设输入方法:仅影像图像、影像所见联合临床病史、图像与文本信息整合。
  • 每例独立评估3次,使用标准化固定提示词。
  • 评估指标为良恶性肿瘤的最终诊断准确率和鉴别诊断准确率。

诊断标准与关键数值

  • 方法一(仅图像):最终诊断准确率总体32%,良性肿瘤25%;鉴别诊断准确率总体54%,良性肿瘤37%。
  • 方法二(影像所见联合病史):最终诊断准确率总体76%,鉴别诊断准确率总体83%。
  • 方法三(图像与文本整合)较方法二的提升无统计学显著性。

临床应用与阅片要点

  • 日常阅片报告中应提供结构化、标准化的影像所见描述,以利于大语言模型辅助诊断工具的发挥。
  • 不应将GPT-4o仅凭图像输入的结果作为诊断依据,需结合完整临床和影像文本信息。
  • 使用前应了解模型对良性肿瘤的诊断准确率偏低,需警惕假阴性。

局限性

  • 回顾性单队列研究,结论临床普适性有限。
  • 原文指出需进行前瞻性验证后才能推进更广泛的临床应用。

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: This study aimed to retrospectively evaluate the diagnostic performance and reliability of GPT-4o, a Large Language Model (LLM), in interpreting radiological images, imaging findings, and patient clinical histories for brain tumor diagnosis within a pre-operative, pathology-confirmed patient cohort. METHODS: We retrospectively analyzed 239 pre-operative brain tumor cases diagnosed between December 2022 and September 2024. GPT-4o’s diagnostic accuracy was assessed under three predefined input methods: radiological images alone (Method 1), radiological findings combined with patient clinical histories (Method 2), and integration of both images and textual information (Method 3). Each case was evaluated independently three times using standardized, fixed prompts. Diagnostic performance was evaluated based on final and differential diagnosis accuracy rates for benign and malignant tumors. RESULTS: Method 1 yielded relatively low final diagnostic accuracies (32% overall, 25% benign tumors) and moderate differential diagnostic accuracies (54% overall, 37% benign tumors). Method 2 demonstrated significantly higher accuracies, reaching 76% for final diagnosis (overall group) and 83% for differential diagnosis (overall group). Method 3 provided slight improvements compared to Method 2, although differences were not statistically significant. CONCLUSION: In this pre-operative brain tumor cohort, GPT-4o demonstrated higher diagnostic accuracy when provided with structured radiological findings and patient histories compared to image analysis alone. These results suggest the potential value of standardized textual inputs to enhance GPT-4o’s diagnostic performance. However, the clinical generalizability of these findings remains limited, and further prospective validation is needed before broader clinical adoption.

来源:Neuroradiology,查看原文。版权归原作者所有。

引用

  1. Sun, Xie, Wang, et al. Integrating large language models (LLMs) into neuro-oncologic radiology: a retrospective feasibility assessment of GPT-4o for brain tumor diagnosis. Neuroradiology. 2026. PMID: 41549106. doi:10.1007/s00234-025-03896-4

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。