【摘要】 目的 探讨人工智能（AI）辅助系统在宫颈薄层液基细胞学研究中的应用。方法 收集2022年7月至2023年2月湖州市妇幼保健院的1 994例宫颈薄层液基细胞学涂片样本。以阴道镜活检结果宫颈上皮内病变（CIN）1级及以上（≥CIN 1级）为金标准，分别采用人工阅片、AI阅片和AI辅助系统阅片（AI辅助阅片）3种方式对宫颈细胞学液基薄层涂片进行阅片。比较不同宫颈薄层液基细胞学检测（TCT）结果下不同年龄段样本3种阅片方式阳性率、灵敏度和特异度以及3种阅片方式的阅片时长。结果 1 994例样本中，活检结果≥CIN 1级共102例，总体阳性率为5.12%,TCT结果为意义不明的非典型鳞状细胞（ASCUS）及以上（≥ASC-US）的阳性率为9.63%，低度鳞状上皮内病变（LSIL）及以上（≥LSIL）的阳性率为4.91%。TCT结果≥ASC-US时，≥30岁组样本AI阅片阳性率高于18～29岁组，差异有统计学意义（P<0.05）；而18～29岁和30岁组人工阅片和AI辅助阅片阳性率比较差异均无统计学意义（均P>0.05）。TCT结果≥LSIL时，18～29岁和30岁组人工阅片、AI阅片和AI辅助阅片阳性率比较差异均无统计学意义（均P>0.05）。TCT结果≥ASC-US时，人工阅片的灵敏度及特异度分别为100.00%、95.30%,AI辅助阅片的灵敏度及特异度分别为95.10%、94.66%，两种阅片方式的灵敏度、特异度比较差异均无统计学意义（均P>0.05）。TCT结果≥LSIL时，人工阅片的灵敏度及特异度分别为66.67%、98.78%,AI辅助阅片的灵敏度及特异度分别为82.35%、97.15%,AI辅助阅片的灵敏度高于人工阅片，差异有统计学意义（P<0.05），而特异度比较差异无统计学意义（P>0.05）。与人工阅片比较，AI辅助阅片的总阅片时长减少25.34 h，总阅片时长减少率为50.76%(25.34/49.92)；平均每张涂片减少45.75 s，平均阅片时长减少率为50.77%(45.75/90.12)。结论 病理医师运用AI辅助系统进行宫颈癌变筛查时，可以保证阅片准确度，提高细胞学阅片效率。更多还原

【Abstract】 Objective To explore the application of artificial intelligence(AI) assisted diagnostic systems in cervical thinprep fluid-based cytology research. Methods Cervical thinprep fluid-based cytological smear samples of 1 994 cases in Huzhou Maternal & Child Health Care Hospital were collected from July 2022 to February 2023. With grades 1 or above in vaginal biopsy results of cervical intraepithelial lesions(≥ CIN 1+) as the gold standard, three methods were used to read cervical thinprep fluid-based cytology smears: manual reading, artificial AI reading, and AI assisted system reading(AI assisted reading). Statistical methods were used to compare the differences in the positive rates detected by the three reading methods among different age groups, and to compare the sensitivity, specificity, and time needed for reading among the three reading methods, so as to explore the diagnostic efficiency of the AI assisted system in cervical fluid-based cytology research. Results Out of 1 994 samples, 102 showed ≥ CIN 1+ results in colposcopy biopsy, with an overall positive rate of 5.12%. The positive rate for atypical squamous cells of undetermined significance and above(≥ ASC-US) in cervical thinprep fluid-based cytologic test(TCT) was 9.63%, while that for low-grade squamous intraepithelial lesion and above(≥ LSIL) in TCT was 4.91%. When the TCT results were ≥ ASC-US, the positive rate by AI reading among patients aged ≥ 30 years was significantly higher than that of the 18-29 age groups(P<0.05), while the positive rate by manual reading or AI assisted reading showed no differences between the two age groups(both P>0.05). When the TCT results were ≥ LSIL, the positive rate via the three reading methods showed no statistical differences between age groups of 18-29 and ≥ 30 years(all P>0.05). When the TCT results were ≥ ASC-US, the sensitivity and specificity of manual reading were100.00% and 95.30%, respectively, while those of AI assisted reading were 95.10% and 94.66%, respectively. No statistical differences were observed in the sensitivity and specificity of the two reading methods(both P>0.05). When TCT results were ≥ LSIL, the sensitivity and specificity of manual reading were 66.67% and 98.78%, respectively, while those of AI assisted reading were 82.35% and 97.15%, respectively. A statistical difference in sensitivity was observed between the two methods of reading(P<0.05), while there was no statistical difference in specificity(P>0.05). Compared with manual reading, the total reading time of AI assisted reading was reduced by 25.34 hours, and 50.76%(25.34/49.92); the average length of each smear decreased by 45.75 s, and the average reading time decreased by 50.77%(45.75/90.12).Conclusion Pathologists can use AI assisted systems to screen for cervical cancer, ensuring the accuracy of reading and improving the efficiency of cytological reading.更多还原