Adaptability and limitations of an artificial intelligence-assisted cervical cytology model in real-world primary care screening

YU Sunxing; XU Zhenguo; JIANG Feng

doi:10.3969/j.issn.1007-8096.2025.12.004

Abstract：Objective To evaluate the adaptability and limitations of an artificial intelligence(AI)-assisted thin-prep cytology test(TCT)model in real-world screening settings of primary care institutions,and to explore its performance across different lesion types,screening processes,and physician acceptance,aiming to provide evidence for optimizing cervical cancer screening strategies.Methods A prospective controlled study was conducted,enrolling 1,200 women undergoing cervical cancer screening at the Second Affiliated Hospital of Fujian University of Traditional Chinese Medicine and its affiliated community hospitals.Participants were assigned to three interpretation groups:AI-assisted interpretation,manual interpretation,and AI-human collaborative interpretation.Using HPV testing and histopathological examination as reference standards,the sensitivity,specificity,negative predictive value(NPV),interpretation time,and misinterpretation types of the three groups in identifying≥CIN2 lesions were assessed.Structured questionnaires and interviews with physicians were conducted to evaluate their subjective experiences regarding system usability,interpretation reliability,and operational interference.Results The AI model demonstrated a sensitivity of 92.8%,specificity of 86.5%,and NPV of 97.6%in identifying≥CIN2 lesions.It effectively pre-screened negative slides,with an average interpretation time of 11.5 seconds.The manual interpretation group(Group A)showed a sensitivity of 91.2%,specificity of 89.3%,NPV of 96.4%,and an average interpretation time of 94.2 seconds.The AI-human collaborative group achieved the highest sensitivity(95.3%)and the lowest false-negative rate.However,the AI model exhibited a tendency for misinterpretation in cases of atypical glandular cells(AGC),mixed lesions,and rare cell types.Physician feedback indicated that 82%believed AI improved work efficiency,78%acknowledged its value in alerting for low-grade lesions,yet 56%still preferred manual review for borderline cases.Conclusion The AI-assisted cervical cytology model demonstrates excellent efficiency and preliminary screening capability in real-world primary care settings,particularly suitable for rapid triage and negative case exclusion in large-scale screenings.While AI is effective for initial rapid triage in primary care,complex lesions require manual review.The AI-human collaborative approach represents a feasible strategy for improving screening quality.