诊断病理学杂志

目的 探讨人工智能（AI）辅助诊断在宫颈液基细胞学筛查中的应用价值。方法 收集常州市妇幼保健院2024年1月至2025年8月常州市宫颈癌筛查中的97 201例宫颈细胞学样本，液基薄层细胞学制片后将玻片利用扫描仪进行数字化呈现，采用AI辅助诊断进行识别及分析，以2022版宫颈细胞学TBS报告分类系统为评判标准,给出具体的诊断。将非典型宫颈鳞状上皮细胞/腺上皮及以上级别作为阳性标准。评估AI辅助诊断以及人工阅片方式在宫颈细胞学筛查的敏感度及特异度，分析两种阅片方式筛查的敏感度、特异度、假阴性率及假阳性率。结果 AI辅助诊断在本次大规模人群宫颈癌筛查中的成功率为99.91%(97 114/97 201)。AI辅助诊断与人工阅片相比，在NILM、ASCUS、ASC-H、LISL、HSIL、SCC、腺上皮异常的诊断一致率分别为98.99%(8 913/9 004)、84.32%(3 751/4 449)、87.58%(213/244)、80.58%(1 282/1 591)、97.06%(324/334)、100%(6/6)、68.59%(142/207)，二者一致性较好（Kappa系数>0.800）。AI辅助诊断在宫颈癌筛查中的灵敏度、特异度、阳性预测值、阴性预测值分别为94.51%、99.10%、87.42%、99.63%，作为宫颈癌筛查工具效果较好。结论 AI辅助诊断在宫颈癌筛查中与人工阅片的一致性良好，AI辅助诊断协同人工复核可有效提高宫颈液基细胞学筛查的敏感性和特异性，降低漏诊风险，极大的提高了阅片效率，在面对大规模人群筛查中，AI辅助诊断系统联合人工复核将阅片效率提高了3～6倍，在工作效率提升上具有显著的优势。

Objective To explore the application value of artificial intelligence(AI)-assisted diagnosis in cervical liquid-based cytology screening. Methods A total of 97,201 cervical cytology samples were collected from the cervical cancer screening program conducted at Changzhou Maternal and Child Health Hospital from January 2024 to August 2025. After preparing liquid-based thin-layer cytology slides, the slides were digitally scanned and analyzed using an AI-assisted diagnostic system. The 2022 version of The Bethesda System(TBS) for cervical cytology reporting was adopted as the diagnostic standard. Atypical squamous cells of undetermined significance(ASC-US) and higher-grade lesions were defined as positive. The sensitivity, specificity, false-negative rate, and false-positive rate of AI-assisted diagnosis and manual screening were evaluated. Results The success rate of AI-assisted diagnosis in this large-scale cervical cancer screening was 99.91%(97,114/97,201). Compared with manual screening, AI demonstrated strong diagnostic agreement across categories: 98.99%(8,913/9,004) in NILM, 84.32%(3,751/4,449) in ASCUS, 87.58%(213/244) in ASC-H, 80.58%(1,282/1,591) in LSIL, 97.06%(324/334) in HISL, 100%(6/6) in SCC, and 68.59%(142/207) in glandular abnormalities. The overall concordance was excellent(Kappa coefficient >0.800). For cervical cancer screening, AI-assisted diagnosis achieved a sensitivity of 94.51%, specificity of 99.10%, positive predictive value(PPV) of 87.42%, and negative predictive value(NPV) of 99.63%, indicating good performance as a screening tool. Conclusion AI-assisted diagnosis demonstrates strong concordance with manual screening in cervical cancer detection. The combination of AI-assisted diagnosis and manual review significantly improves the sensitivity and specificity of cervical liquid-based cytology screening, reduces the risk of missed diagnoses, and greatly enhances diagnostic efficiency. In large-scale cervical cancer screening programs, the combination of AI-assisted diagnostic system and manual review has improved slide-reading efficiency by 3 to 6 times, demonstrating significant advantages in workflow optimization.