基于大语言模型与视觉Transformer的AI驱动颅内易损动脉粥样硬化斑块诊断:一项多中心研究
本回顾性多中心研究纳入726例症状性颅内动脉粥样硬化狭窄患者的1806个斑块,评估大语言模型(ChatGPT-4o、DeepSeek-V3、Moonshot AI)与深度学习模型(ResNet50、ViT)基于高分辨率血管壁成像诊断易损斑块的效能。ChatGPT-4o诊断AUC最高(0.874),ViT优于ResNet50(AUC 0.913 vs 0.845)。LLMs与ViT联合显著提高初级医生诊断准确率并将斑块评估时间从301秒缩短至174秒(p<0.05)。
要点
- ViT分类易损斑块的AUC为0.913,优于ResNet50的0.845。
- ChatGPT-4o在LLMs中诊断性能最高,AUC为0.874。
- LLMs与ViT联合使初级医生诊断准确率显著提高,斑块评估时间从301秒降至174秒(p<0.05)。
核心结论
- 整合大语言模型与深度学习模型可提升诊断性能与效率,尤其对初级医生;该方法可能通过提高诊断准确率并缩短斑块分析时间,减轻医疗系统负担,尤其在资源有限环境中。
研究设计与样本
- 回顾性研究,纳入2018年6月至2024年6月症状性颅内动脉粥样硬化狭窄患者726例,共分析1806个斑块;使用高分辨率血管壁成像(HR-VWI)报告训练LLMs(ChatGPT-4o、DeepSeek-V3、Moonshot AI)提取诊断见解,并使用ResNet50与Vision Transformer(ViT)对易损斑块进行分类;由初级和高级医生评估诊断准确率、敏感度、特异度及时间效率。
诊断标准与关键数值
- ChatGPT-4o诊断性能最高,AUC为0.874。
- ViT分类易损斑块的AUC为0.913,优于ResNet50的AUC 0.845。
- LLMs与ViT联合使初级医生诊断准确率显著提高,斑块评估时间从301秒缩短至174秒(p<0.05)。
临床应用与阅片要点
- 作者认为,LLMs与深度学习模型的整合可提升诊断性能与效率,尤其对初级医生,并可能减轻医疗系统负担,尤其在资源有限环境中。
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
OBJECTIVES: High-resolution vessel wall imaging (HR-VWI) is essential for diagnosing vulnerable intracranial atherosclerotic plaques, but its interpretation requires expertise. This study investigates the integration of large language models (LLMs) and deep learning (DL) for more efficient diagnosis.
MATERIALS AND METHODS: A retrospective study of symptomatic intracranial atherosclerotic stenosis patients (June 2018-June 2024) was conducted. LLMs (ChatGPT-4o, DeepSeek-V3, and Moonshot AI) were trained on HR-VWI reports to extract diagnostic insights. Additionally, DL models, ResNet50 and Vision Transformer (ViT), were used to classify vulnerable plaques. Diagnostic accuracy, sensitivity, specificity, and time efficiency were evaluated with both junior and senior doctors.
RESULTS: A total of 1806 plaques from 726 patients were analyzed. ChatGPT-4o exhibited the highest diagnostic performance (AUC: 0.874). Among DL models, ViT outperformed ResNet50 (AUC: 0.913 vs. 0.845). LLMs and ViT significantly improved junior doctors' diagnostic accuracy and reduced plaque assessment time (from 301 s to 174 s, p < 0.05).
CONCLUSION: The integration of LLMs and DL models enhanced diagnostic performance and efficiency, especially for junior doctors. This approach could reduce the burden on healthcare systems, particularly in resource-limited settings, by improving diagnostic accuracy and reducing the time required for plaque analysis.
KEY POINTS: Question How can AI models assist less experienced doctors in accurately identifying vulnerable intracranial plaques on high-resolution vessel wall imaging (HR-VWI) and improve diagnostic efficiency? Findings The integration of large language models (LLMs) and deep learning (ViT) significantly improves diagnostic accuracy and efficiency, reducing assessment time for vulnerable plaques. Clinical relevance This study demonstrates that combining LLMs and deep learning enables junior doctors to achieve near-expert accuracy in diagnosing vulnerable plaques, potentially reducing stroke risk and easing diagnostic burdens in resource-limited healthcare environments.
引用
- Li ZA, Gao Y, Ji K, et al. AI-driven diagnosis of vulnerable intracranial atherosclerotic plaques using large language models and vision transformers: a multi-center study. Eur Radiol. 2025. PMID: 41123642. doi:10.1007/s00330-025-12065-3
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。