人工智能用于肺炎诊断：多模态学习、融合策略与可解释性的系统综述。
完整摘要 · 中文译文
肺炎仍然是全球重大公共卫生挑战，尤其是在儿童、老年人以及免疫系统受损的个体中。尽管深度学习在自动化肺炎诊断方面取得了巨大进展，但当前大多数方法是单模态的，仅使用胸部X光图像。这导致其鲁棒性、跨机构泛化能力和临床适用性有限。此外，现有综述主要侧重于总结单一数据模态或模型架构，缺乏对多模态融合策略、可解释性方法、基准数据集和新型人工智能范式的比较分析。为填补这些空白，本系统综述提供了一个通用分析框架，用于系统比较用于肺炎诊断的人工智能领域中的多模态深度学习架构、融合策略、可解释性方法、基准数据集和新兴技术。本文对研究文章进行了系统综述，以识别最新进展、技术挑战和未来研究方向。证据表明，多模态模型可以提供来自不同来源的信息，并可能在特定场景下提高诊断性能。然而，研究之间的直接比较因数据集、临床任务和验证方案的不同而受到限制。基于注意力和混合融合策略可以从多种模态中学习有用信息。可解释人工智能方法提供了提高模型预测的透明度和可解释性的机制。 然而，由于缺乏大规模多模态数据集、数据异质性、模态缺失、评估方案不一致以及缺乏外部验证，临床转化仍受到限制。总之，本综述总结了当前的研究趋势，概述了重要的技术挑战，并为未来构建强大、可解释且临床可行的人工智能系统以用于下一代肺炎诊断提供了研究议程。
完整摘要 · 英文原文
Pneumonia remains a major global public health challenge, particularly among children, older adults, and individuals with weakened immune systems. Although deep learning has made great progress in automated pneumonia diagnosis, most current methods are unimodal, using only chest X-ray images. This results in limited robustness, cross-institutional generalization and clinical applicability. Moreover, the current reviews mainly focus on summarising single data modalities or model architectures and lack comparative analysis of multimodal fusion strategies, explainability methods, benchmark datasets, and new artificial intelligence paradigms. To fill these gaps, this systematic review offers a common analytical framework for systematically comparing multimodal deep learning architectures, fusion strategies, explainability methods, benchmark datasets, and emerging technologies in the field of artificial intelligence for pneumonia diagnosis. Research articles were systematically reviewed to identify the latest developments, technological challenges, and future research directions. Evidence suggests that multimodal models can provide information from different sources and may improve diagnostic performance in specific settings. However, direct comparisons between studies are limited by differences in datasets, clinical tasks, and validation protocols. Attention-based and hybrid fusion strategies can learn useful information from various modalities. Explainable artificial intelligence methods provide mechanisms for improving the transparency and interpretability of model predictions.The clinical translation is still limited, however, due to the lack of large-scale multimodal datasets, data heterogeneity, missing modalities, inconsistent evaluation protocols, and lack of external validation. In summary, this review summarizes the current research trends, outlines important technological challenges, and provides a research agenda for the future to build strong, interpretable, and clinically viable artificial intelligence systems for next-generation pneumonia diagnosis.