人工智能在宫颈疾病诊断中的应用进展

Progress in the Application of Artificial Intelligence in the Diagnosis of Cervical Diseases

摘要: 宫颈癌是全球女性常见恶性肿瘤之一，其筛查与癌前病变的早期识别对降低疾病负担具有重要意义。宫颈细胞学与阴道镜检查作为宫颈癌筛查体系中的关键环节，在临床分流和风险评估中发挥着重要作用，但二者均不同程度依赖阅片经验与主观判断，存在诊断一致性不足、质量控制难度较大及跨中心推广受限等现实问题。近年来，人工智能学习与深度学习技术在宫颈细胞学图像分析和阴道镜图像判读中的应用快速发展，为宫颈病变的标准化识别、风险分层及临床辅助决策提供了新的技术路径。本文围绕人工智能在宫颈细胞学与阴道镜中的应用进展，重点总结宫颈鳞状上皮低级别与高级别病变鉴别难点中的研究现状、模型类型及潜在临床价值。现有研究表明，人工智能在异常细胞检测、病变区域定位、阴道镜图像分类及CIN2+/HSIL风险预测等方面表现出较好的应用前景，并呈现出由局部识别向全视野分析、由单模态分类向多模态融合、由单纯追求性能指标向强调可解释性与临床可用性转变的发展趋势。然而，不同研究在数据来源、图像采集条件、标注方式、金标准设定、终点定义及验证策略等方面仍存在较大异质性，限制了模型的泛化能力与真实世界转化。未来研究需进一步加强标准化数据体系建设，推进多中心、跨设备、前瞻性外部验证，促进多模态信息整合与人机协同决策，以推动人工智能在宫颈癌筛查与癌前病变管理中的规范化、可靠化和临床化应用。

Abstract: Cervical cancer is one of the most common malignant tumors among women worldwide. The screening and early identification of precancerous lesions are of great significance in reducing the disease burden. Cervical cytology and colposcopy, as key components of the cervical cancer screening system, play an important role in clinical triage and risk assessment. However, both rely to varying degrees on radiographic experience and subjective judgment, leading to practical issues such as insufficient diagnostic consistency, difficulty in quality control, and limited cross-center promotion. In recent years, the rapid development of artificial intelligence learning and deep learning technology in cervical cytology image analysis and colposcopy image interpretation has provided a new technological path for standardized identification, risk stratification, and clinical auxiliary decision-making of cervical lesions. This article focuses on the application progress of artificial intelligence (AI) in cervical cytology and colposcopy, emphasizing the current research status, model types, and potential clinical values in the differentiation of low-grade and high-grade squamous intraepithelial lesions. Existing research indicates that AI demonstrates promising application prospects in abnormal cell detection, lesion area localization, colposcopy image classification, and CIN2+/HSIL risk prediction. It also presents a development trend from local recognition to full-field analysis, from single-modality classification to multi-modality fusion, and from solely pursuing performance metrics to emphasizing interpretability and clinical usability. However, significant heterogeneity exists in different studies regarding data sources, image acquisition conditions, annotation methods, gold standard setting, endpoint definition, and validation strategies, limiting the generalization ability and real-world translation of models. Future research needs to further strengthen the construction of standardized data systems, promote multi-center, cross-device, prospective external validation, facilitate multi-modality information integration and human-machine collaborative decision-making, and drive the standardized, reliable, and clinical application of AI in cervical cancer screening and precancerous lesion management.