人工智能驱动的肾脏病理学用于肾小球疾病分类：系统综述与荟萃分析。
完整摘要 · 中文译文
背景：诊断性能在不同肾小球疾病（GD）亚型之间存在差异，人工智能（AI）是否优于不同经验水平的病理学家仍不确定。 目的：评估基于病理学的AI模型用于GD分类，并将其性能与病理学家进行比较。 方法：检索PubMed、Embase、Web of Science和Cochrane Library，截至2026年7月15日。纳入使用病理图像并以病理诊断作为参考标准的研究。采用随机效应模型合并敏感性、精确率、准确率、F1评分和曲线下面积（AUC）。 结果：共纳入15项研究，包含39,536个验证样本单元，这些单元不一定是唯一患者。对于至少有10个验证数据集的亚型，AI在膜性肾病（MN；敏感性0.96，精确率0.94，准确率0.96，F1评分0.95，AUC 0.98）、IgA肾病（IgAN；敏感性0.92，精确率0.91，准确率0.94，F1评分0.90，AUC 0.96）和微小病变病（MCD；敏感性0.92，精确率0.87，准确率0.96，F1评分0.89，AUC 1.00）方面取得了高性能。AI在IgAN、MN和MCD方面也显示出高于高级病理学家的准确率；然而，对照证据稀少，应谨慎解读。大多数纳入研究为回顾性研究，并且在数据集、成像模式、模型架构和验证策略之间观察到显著异质性。 结论：基于病理学的AI在GD分类方面显示出强大潜力，但当前的头对头证据不足以确立优越性 优先于病理学家，尤其是资深病理学家。需要整合多模态临床数据并开展标准化外部验证的前瞻性多中心研究。 已知内容：正在开发基于肾脏病理图像的人工智能，以支持肾小球疾病的分类。 本研究补充的内容：本系统综述综合了病理图像人工智能模型，并评估了其性能与验证背景及异质性的关系。潜在影响：这些模型可在病理学家监督下支持可重复的筛查和定量评估，但仍需前瞻性多中心验证。
完整摘要 · 英文原文
BACKGROUND: Diagnostic performance varies across glomerular disease (GD) subtypes, and whether artificial intelligence (AI) outperforms pathologists with different experience levels remains uncertain. OBJECTIVE: To evaluate pathology-based AI models for GD classification and compare their performance with pathologists. METHODS: PubMed, Embase, Web of Science, and Cochrane Library were searched through 15 July 2026. Studies using pathology images and pathology diagnosis as the reference standard were included. Random-effects models pooled sensitivity, precision, accuracy, F1 score and area under the curve (AUC). RESULTS: Fifteen studies comprising 39,536 validation sample units, not necessarily unique patients, were included. For subtypes with at least 10 validation datasets, AI achieved high performance for membranous nephropathy (MN; sensitivity 0.96, precision 0.94, accuracy 0.96, F1 score 0.95, AUC 0.98), IgA nephropathy (IgAN; sensitivity 0.92, precision 0.91, accuracy 0.94, F1 score 0.90, AUC 0.96), and minimal change disease (MCD; sensitivity 0.92, precision 0.87, accuracy 0.96, F1 score 0.89, AUC 1.00). AI also showed higher accuracy than senior pathologists for IgAN, MN, and MCD; however, comparator evidence was sparse and should be interpreted cautiously. Most included studies were retrospective, and substantial heterogeneity was observed across datasets, imaging modalities, model architectures, and validation strategies. CONCLUSIONS: Pathology-based AI shows strong potential for GD classification, but current head-to-head evidence is insufficient to establish superiority over pathologists, particularly senior pathologists. Prospective multicenter studies integrating multimodal clinical data and standardized external validation are needed. WHAT IS KNOWN: Artificial intelligence based on renal pathology images is being developed to support classification of glomerular disease. WHAT THIS STUDY ADDS: This systematic review synthesizes pathology-image AI models and evaluates performance in relation to validation context and heterogeneity.Potential impact: These models may support reproducible screening and quantitative assessment under pathologist oversight, but prospective multicenter validation remains necessary.