摘要： 目的 探讨人工智能（AI）辅助细胞学系统对宫颈细胞学初筛及Bethesad系统（TBS）分级诊断的效能，明确AI在宫颈病变筛查中的灵敏度及阴性诊断可靠性。方法 选取我院29 354例宫颈细胞学涂片进行AI及人工阅片诊断，其中2 246例取得病理活检结果，比较分析AI的灵敏度和特异度。结果 以人工复核结果为液基细胞学检查（TCT）最终诊断，AI诊断的灵敏度为98.59%，特异度为60.41%，阴性预测值高达99.66%。AI灵敏度高，可减少漏诊；特异度较低，阳性病例需人工复核； AI阴性诊断准确性极高，无需人工复核。AI漏诊率（1.41%）低于人工（2.69%），差异有统计学意义，无低度鳞状上皮内病变（LSIL）及以上级别漏诊。AI与人工诊断TBS系统总体符合率为57.71%，Kappa值为0.371，提示AI分级诊断准确性低。以组织病理学为金标准，AI灵敏度为97.15%、特异度为38.01%、阴性预测值为83.97%。人工阅片灵敏度为89.64%、特异度为48.26%、阴性预测值为64.69%。提示AI有较高的灵敏度及阴性预测值。结论 AI是一种有效的宫颈病变筛查的新手段，灵敏度高，漏诊率低，可有效做到阴、阳性分流，适用于宫颈病变的大规模筛查。

人工智能在宫颈病变初步筛查中的应用

张舒婉, 张文川, 张喆, 迂金洋, 谢伦昀, 吕庆杰

中国医科大学附属盛京医院病理科, 沈阳 110004

收稿日期:2022-11-16出版日期:2023-08-30发布日期:2023-08-07
通讯作者:吕庆杰E-mail:lvqjie@163.com
作者简介:张舒婉(1994-),女,医师,硕士.
基金资助:国家自然科学基金（82072095）；盛京医院345人才工程


关键词: 人工智能, 宫颈细胞学, 筛查
Abstract: Objective This study aimed to explore the efficacy of artificial intelligence (AI) assisted cytology system in the primary screening diagnosis of cervical cytology,assess the efficacy of AI in grading diagnosis of cervical cytology TBS,and evaluate the sensitivity of AI in screening cervical lesions and the reliability of negative diagnosis. Methods 29 354 cervical cytology smear specimens in our hospital were diagnosed using AI and by manual reading. Overall,2 246 of these specimens were pathological biopsies. The sensitivity and specificity of AI were compared with those of manual reading and were analyzed. Results The final diagnosis of TCT was based on the results of the manual review. The sensitivity,specificity,and negative predictive value of AI diagnoses were 98.59%,60.41%,and 99.66%,respectively. Our findings revealed the following:AI has high sensitivity,which can reduce the rate of of missed diagnoses; the specificity of AI is low,and positive cases need manual review; and the accuracy of AI in making negative diagnoses is extremely high,and manual review is not needed. The missed diagnosis rate of AI (1.41%) was lower than that of manual diagnosis (2.69%),and the difference was statistically significant. There was no missed diagnosis of LSIL or above. The overall coincidence rate between AI and manual diagnosis TBS system was 57.71% and the Kappa value was 0.371,which indicated that the accuracy of AI grading diagnosis was low. Taking histopathology as the gold standard,the sensitivity,specificity,and negative predictive value of AI were 97.15%,38.01%,and 83.97%,respectively. Furthermore,the sensitivity,specificity,and negative predictive value of manual reading were 89.64%,48.26%,and 64.69%,respectively. AI has high sensitivity and negative predictive value. Conclusion AI is an effective novel method for screening cervical lesions with high sensitivity and low rates of missed diagnoses. AI can effectively distinguish negative cases from positive cases and is suitable for large-scale screening of cervical lesions.
Key words: artificial intelligence, cervical cytology, screening
PDF全文下载地址:

https://journal.cmu.edu.cn/CN/article/downloadArticleFile.do?attachType=PDF&id=3262