Comparative study on quality control models for cervical liquid-based thin-layer cytology smears constructed using artificial intelligence techniques
- VernacularTitle:人工智能构建宫颈液基薄层细胞涂片质控模型对比研究
- Author:
Yongqin WEN
1
;
Ruoyu ZHANG
1
;
Xianlei LI
1
;
Hua XU
1
;
Xiaomin LIAO
1
;
Wei YUAN
1
;
Weibiao YE
1
Author Information
- Publication Type:Journal Article
- Keywords: artificial intelligence(AI)-assisted analysis; cervical cytology; cervical cancer screening
- From: Journal of Xi'an Jiaotong University(Medical Sciences) 2025;46(3):544-550
- CountryChina
- Language:Chinese
- Abstract: Objective To construct a quality control model for cervical liquid-based thin cell smears using two different artificial intelligence(AI)techniques and to compare the total use of the two methods to improve the level of quality control of cervical liquid-based thin cell smears through the assistance of hybrid AI.Methods In this study,105 cervical liquid-based thin cell smear samples were used.Convolutional neural network(CNN)algorithm and Transformer network algorithm were used as specific AI algorithms in the AI model.The labeled features included the number of cells in the slice,excessive red blood cells,excessive inflammatory cells,and air bubbles.The smear samples were pre-processed and digitized by smear,followed by image segmentation and feature extraction.Using the labeled feature data,machine learning models were trained and optimized.Statistical AI and physician QC results were analyzed by calculating KAPPA index,sensitivity,specificity,area under the curve(AUC),and other indexes for AI QC results.Results CNN algorithm QC results in normal smear,inflammatory background and bloody background were significantly different from the expert review QC results(P＜0.001).Transformer algorithm QC results were similar to the expert review results,with no statistical difference(P＞0.05).General practitioner QC results were statistically different from the expert review QC results in normal smear detection rate and bloody background(P＜0.001).CNN algorithm Kappa value was 0.567,which had medium consistency with expert review results.Transformer algorithm Kappa value was 0.890,with the best consistency with expert review results.General practitioner Kappa value was 0.675,which had better consistency with expert review results.Using the expert review results as a reference standard,the predictive efficacy of the Transformer algorithm and the general practitioners' QC results was evaluated,and the predictive efficacy of the Transformer algorithm was higher than that of the general practitioners in detecting hemorrhagic backgrounds and normal smears(inflammatory backgrounds:AUC=1.000;normal smears:AUC=0.768)(hemorrhagic backgrounds:AUC=0.849;normal smears:AUC=0.849;normal smear:AUC=0.500).Conclusion In this study,we found that the Transformer algorithm was effective in improving the quality control of cervical liquid-based thin-layer cell smears by assisting doctors to perform smear quality control scoring and improving the efficiency and accuracy of smear sample quality control.It can be used as a new quality control method for cervical cancer cytological screening and has potential clinical applications.
