Artificial intelligence-assisted analysis in cervical liquid-based cytology:application and advantages and limita-tions

GUO Yu-juan; XIE Nan-yu; ZHAO Pei-yao; PENG Hai-yan; YAN Shan-shan

doi:10.13820/j.cnki.gdyx.20252173

Abstract：Objective To explore the clinical application value of artificial intelligence(AI)assisted analysis system in cervical liquid-based cytology(LBC)and analyze its advantages and limitations.Methods The AI-assis-ted system analyzed 19 057 LBC samples collected from patients at Guangdong Women and Children's Hospital between November 10,2023,and May 20,2024.Cytological interpretations included cervical intraepithelial lesions and microbial infections.AI results were compared with manual reading,pathological biopsy,and HPV testing.Results AI reading showed a total positive detection rate of 41.32％,significantly higher than manual reading(10.50％),especially in ASC-US and ASC-H categories,though with higher false-positive rates.AI's negative predictive value for LSIL and higher-grade lesions was high.Both AI and manual reading showed high sensitivity for CIN1+cases,but manual read-ing had slightly higher sensitivity.For HSIL+cases,manual reading specificity was higher.HPV testing showed 92.55％positivity among positive samples,with HPV52 most common,followed by 16,58,51,53,56,and 39.AI sensitivity for microbial detection was low overall,slightly better for bacterial vaginosis,but poor for Candida and Trichomonas.AI also had blind spots in glandular epithelial lesion detection.Conclusion AI-assisted LBC reading has high potential in rul-ing out negative cases,reducing cytologists' workload and improving efficiency,but misdiagnosis and missed diagnosis re-main issues.AI is limited in detecting microbial infections and glandular lesions,requiring further deep learning improve-ment.Wider application and refinement of AI systems could significantly advance cervical cancer screening.