基于神经影像的人工智能诊断额颞叶痴呆:系统综述与荟萃分析
该系统综述与荟萃分析纳入23项研究共3670例受试者,评估基于神经影像的AI模型区分额颞叶痴呆(FTD)及其主要临床亚型与认知正常对照的诊断准确性。汇总敏感度0.91(95% CI 0.87-0.93),特异度0.92(95% CI 0.89-0.94),AUC 0.96(95% CI 0.94-0.98),语义变异型原发性进行性失语(svPPA)亚型AUC最高达0.98(95% CI 0.97-0.99)。证据确定性受高偏倚风险和研究异质性限制,需前瞻性多中心验证。
要点
- AI模型鉴别FTD与认知正常对照的汇总AUC达0.96,敏感度0.91、特异度0.92。
- svPPA亚型诊断性能最高,AUC 0.98,优于其他FTD亚型。
- 不同影像模态(sMRI、fMRI、PET、多模态)AUC均稳定在0.94-0.98。
诊断标准与关键数值
- AI模型区分FTD与认知正常对照的汇总敏感度0.91(95% CI 0.87-0.93;I2=73.2%),特异度0.92(95% CI 0.89-0.94;I2=80.2%),AUC 0.96(95% CI 0.94-0.98)。
- svPPA亚型诊断AUC最高,为0.98(95% CI 0.97-0.99)。
- 不同影像模态(sMRI、fMRI、PET、多模态)AUC范围为0.94-0.98。
- 2009至2025年报告的AUC值呈轻微下降趋势,但多数模型AUC仍高于0.90。
临床应用与阅片要点
- AI模型可作为FTD诊断流程的辅助工具,尤其在对svPPA亚型进行鉴别时参考价值较高。
- 当前证据不支持将AI模型作为独立诊断工具,阅片和报告时应结合临床和影像综合判断。
- 注意研究间异质性较高(敏感度I2=73.2%,特异度I2=80.2%),应用时需考虑模型泛化能力。
局限性
- 纳入研究频繁存在高偏倚风险,证据确定性受限。
- 研究间异质性较高(敏感度I2=73.2%,特异度I2=80.2%)。
- 亚组样本量小,亚型分析结果需谨慎解读。
- 缺乏前瞻性多中心外部验证,临床通用性尚未确认。
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
OBJECTIVE: This meta-analysis aimed to evaluate the diagnostic accuracy of neuroimaging-based artificial intelligence (AI) models for detecting frontotemporal dementia (FTD) and its primary clinical subtypes.
MATERIALS AND METHODS: A systematic search of PubMed, Embase, Web of Science, and Cochrane Library was conducted up to October 22, 2025. Studies reporting neuroimaging-based AI model performance for diagnosing FTD variants against cognitively normal (CN) controls were included. Study quality and outcome certainty were assessed using the PROBAST + AI tool and GRADE. A bivariate random-effects model was used to calculate pooled sensitivity, specificity, and the area under the curve (AUC).
RESULTS: Twenty-three studies involving 3670 subjects were included. AI models demonstrated a pooled sensitivity of 0.91 (95% CI: 0.87-0.93; I2 = 73.2%), specificity of 0.92 (95% CI: 0.89-0.94; I2 = 80.2%), and an AUC of 0.96 (95% CI: 0.94-0.98). Subtype analysis revealed the highest performance for semantic variant primary progressive aphasia (svPPA), with an AUC of 0.98 (95% CI: 0.97-0.99). Across neuroimaging modalities (sMRI, fMRI, PET, and multimodal), AUC values remained consistently robust, ranging from 0.94 to 0.98. Notably, a temporal analysis indicated a slight declining trend in reported AUC values from 2009 to 2025, though most models maintained an AUC above 0.90.
CONCLUSIONS: Neuroimaging-based AI models show promising diagnostic performance for FTD, especially for the svPPA subtype, supporting their potential as a clinical adjunct. However, the evidence certainty is limited by frequent high risk of bias and study heterogeneity. Prospective, multicenter validation is required to confirm generalizability and clinical utility.
REGISTRATION NUMBER: CRD420251270006.
KEY POINTS: Question Can neuroimaging-based artificial intelligence accurately distinguish frontotemporal dementia and its major clinical subtypes from cognitively normal controls in published diagnostic studies? Findings Across 23 studies, pooled accuracy was high, but estimates were limited by heterogeneity, small subgroup samples, and frequent validation bias. Clinical relevance AI models may support FTD diagnostic workflows, but current evidence favors cautious adjunctive use rather than stand-alone clinical implementation until prospective external validation improves.
引用
- Li Y, He X, Li N, et al. AI-based diagnosis of frontotemporal dementia using neuroimaging: a systematic review and meta-analysis. Eur Radiol. 2026. PMID: 42965691. doi:10.1007/s00330-026-13016-2
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。