目的 系统综述人工智能（artificial intelligence，AI）与多模态融合技术在肺结节良恶性鉴别中的研究进展，重点阐述影像、临床资料及血液生物标志物整合决策的理论基础、关键技术、临床价值与应用边界，为肺结节精准分层管理及临床转化提供参考。方法 基于肺结节管理相关国际指南、经典风险预测模型、近年 AI 影像学研究、多组学与液体活检研究以及方法学规范文献，从多模态技术评估肺结节的意义、影像数据与临床资料融合、血液标志物/循环肿瘤 DNA（circulating tumor DNA，ctDNA）/循环染色体异常细胞（circulating genetically abnormal cells，CAC）/蛋白组学协同价值、多模态模型的临床预测潜力、“辅助决策”而非“替代决策”的边界，以及当前挑战与解决路径等方面进行归纳、分析与综合。结果 当前肺结节管理仍以结节大小、体积、密度、边缘特征、增长动力学及患者年龄、吸烟史、既往肿瘤史等临床危险因素为基础，并依托 Fleischner 学会、英国胸科学会（British thoracic society，BTS）、肺结节分级及数据系统（lung imaging reporting and data system，Lung-RADS）、美国胸外科协会（American association for thoracic surgery，AATS）等指南及 Mayo、Herder、Brock 等经典模型完成风险分层。然而，在亚厘米结节、亚实性结节、多发结节、炎症背景结节及中间风险结节中，单一影像征象或传统模型的校准能力和临床净获益仍有限。AI，尤其是影像组学（radiomics）、深度学习和多模态机器学习，可从 CT 中提取人眼难以稳定识别的高维特征；与年龄、吸烟状态、肺气肿、既往恶性肿瘤史等临床变量融合后，可提升风险评估的稳定性和重分类能力。同时，游离细胞 DNA （cell-free DNA，cfDNA）/ctDNA 甲基化、片段组学、CAC 及蛋白分类器等液体活检手段，为中间风险结节提供了额外的分子与细胞学证据，有助于减少不必要的侵入性操作，并促进真正高危病灶更早进入精准诊断流程。当前多模态模型的主要价值体现在中间风险结节的二次分层、专病门诊流程优化和多学科团队（multidisciplinary team，MDT）决策支持，但其临床落地仍受数据异质性、外部泛化能力不足、前分析标准化不充分、高维低样本、监管与支付机制等因素制约。结论 肺结节良恶性鉴别正由单模态影像判断向“影像—临床—生物标志物”多模态整合决策转变。AI 在当前阶段最合理的定位仍是辅助决策而非替代决策。未来真正可能改变临床实践的，将是经前瞻性验证、具备可解释性与可审计性、符合指南逻辑并可嵌入肺结节专病中心和 MDT工作流的多模态辅助决策系统。

Objective To systematically review recent advances in artificial intelligence (AI) and multimodal fusion for differentiating benign from malignant pulmonary nodules, with a focus on the theoretical basis, key technologies, clinical utility, and practical boundaries of integrated decision-making based on imaging, clinical data, and blood-based biomarkers. Methods International guidelines for pulmonary nodule management, classic risk prediction models, recent AI-based imaging studies, multiomics and liquid biopsy studies, and methodological consensus documents were reviewed. Evidence was synthesized from six perspectives: the significance of multimodal assessment, integration of imaging and clinical variables, synergistic value of blood biomarkers including ctDNA and circulating genetically abnormal cells (CAC), clinical potential of multimodal models, the boundary between decision support and decision replacement, and current challenges with possible solutions. Results Current pulmonary nodule management still relies primarily on nodule size, volume, density, margin characteristics, growth dynamics, and conventional clinical risk factors such as age, smoking history, and prior malignancy, under the framework of established guidelines and prediction models. However, in subcentimeter nodules, subsolid nodules, multiple nodules, inflammation-related nodules, and intermediate-risk nodules, single-modality imaging features and conventional models remain inadequate in calibration and net clinical benefit. AI-based radiomics, deep learning, and multimodal machine learning can extract high-dimensional CT features beyond human visual recognition and improve risk stratification when combined with clinical variables. Meanwhile, liquid biopsy approaches, including cfDNA/ctDNA methylation, fragmentomics, CAC, and proteomic classifiers, provide additional molecular and cellular evidence for intermediate-risk nodules, thereby helping reduce unnecessary invasive procedures and accelerating precision diagnosis in truly high-risk cases. Nevertheless, real-world implementation remains limited by data heterogeneity, insufficient external validation, lack of assay standardization, high-dimensional low-sample-size issues, and regulatory and reimbursement barriers. Conclusion The differential diagnosis of pulmonary nodules is evolving from single-modality imaging judgment toward multimodal integrated decision-making based on imaging, clinical data, and biomarkers. At the current stage, AI should be positioned as a decision-support tool rather than a decision-replacement tool. Future practice-changing systems will likely be prospectively validated, interpretable, auditable, guideline concordant multimodal platforms that can be seamlessly embedded into pulmonary nodule clinics and multidisciplinary workflows.