人工智能分析系统对需氧菌性阴道炎的判读方法初探

Preliminary Study on the Identification of Aerobic Vaginitis by Artificial Intelligence Analysis System

目的开发基于深度学习的人工智能阴道分泌物分析系统,评估自动化镜检对需氧菌性阴道炎(aerobic vaginitis,AV)临床诊断的准确性。方法选取2020年1月–2021年12月就诊于四川大学华西第二医院妇产科3769例患者的阴道分泌物,以人工镜检结果为对照,通过Python Scikit-learn script开发出能识别含中毒颗粒白细胞和基底旁上皮细胞(parabasal epitheliocytes,PBC)的人工智能自动化分析软件(linear kernel SVM algorithm),并利用乳杆菌和AV常见分离菌的标准菌株重新设置细菌分级参数。以人工镜检结果为对照,得到人工智能判断AV评分中各项目各分值之间的受试者工作特征(receiver operating characteristic,ROC)曲线和cut-off值,从而设定出自动化判读AV的参数,初步建立AV自动化分析评分方法。结果共收集到3769份阴道分泌物标本。人工智能识别AV共有5个参数,每个参数有3种程度。乳杆菌与AV常见菌的直径分界值为1.5μm,乳杆菌的自动化判断参数是长径≥1.5μm∶<1.5μm细菌的比值,分界值是2.5和0.5;白细胞(white blood cell,WBC)的自动化判断参数中,WBC绝对数量的分界值是10~3μL^(-1),WBC/上皮细胞的比值分界值是10;含中毒颗粒白细胞的自动化判断参数是含中毒颗粒WBC/WBC比值,分界值是1%和15%;背景菌落的自动化判断参数是<1.5μm,细菌分界值是5×10^(3)μL^(-1)和3×10^(4)μL^(-1),PBC的自动化判断参数是PBC/上皮细胞的比值,分界值是1%和10%。自动化镜检与人工镜检的一致率为92.5%,200例标本中评分一致的有185例,不一致有15例。结论本研究开发的人工智能AV识别软件,其建立的阴道分泌物AV自动化镜检评分方法,检测结果与人工镜检具有较好的一致性,可较为客观、敏感、高效完成临床检验,并降低人工镜检工作负荷。

Objective To develop an artificial intelligence vaginal secretion analysis system based on deep learning and to evaluate the accuracy of automated microscopy in the clinical diagnosis of aerobic vaginitis(AV).Methods In this study,the vaginal secretion samples of 3769 patients receiving treatment at the Department of Obstetrics and Gynecology,West China Second Hospital,Sichuan University between January 2020 and December 2021 were selected.Using the results of manual microscopy as the control,we developed the linear kernel SVM algorithm,an artificial intelligence(AI)automated analysis software,with Python Scikit-learn script.The AI automated analysis software could identify leucocytes with toxic appearance and parabasal epitheliocytes(PBC).The bacterial grading parameters were reset using standard strains of lactobacillus and AV common isolates.The receiver operating characteristic(ROC)curve analysis was used to determine the cut-off value of AV evaluation results for different scoring items were obtained by using the results of manual microscopy as the control.Then,the parameters of automatic AV identification were determined and the automatic AV analysis scoring method was initially established.Results A total of 3769 vaginal secretion samples were collected.The AI automated analysis system incorporated five parameters and each parameter incorporated three severity scoring levels.We selected 1.5μm as the cut-off value for the diameter between Lactobacillus and common AV bacterial isolates.The automated identification parameter of Lactobacillus was the ratio of bacteria≥1.5μm to those<1.5μm.The cut-off scores were 2.5 and 0.5,In the parameter of white blood cells(WBC),the cut-off value of the absolute number of WBC was 10^(3)μL^(-1) and the cut-off value of WBC-to-epithelial cell ratio was 10.The automated identification parameter of toxic WBC was the ratio of toxic WBC toWBC and the cut-off values were 1%and 15%.The parameter of background flora was bacteria<1.5μm and the cut-off values were 5×10^(3)μL^(-1) and 3×10^(4)μL^(-1).The parameter of the parabasal epitheliocytes was the ratio of PBC to epithelial cells and the cut-off values were 1%and 10%.The agreement rate between the results of automated microscopy and those of manual microscopy was 92.5%.Out of 200 samples,automated microscopy and manual microscopy produced consistent scores for 185 samples,while the results for 15 samples were inconsistent.Conclusion We developed an AI recognition software for AV and established an automated vaginal secretion microscopy scoring system for AV.There was good overall concordance between automated microscopy and manual microscopy.The AI identification software for AV can complete clinical lab examination with rather high objectivity,sensitivity,and efficiency,markedly reducing the workload of manual microscopy.