Application of AI-assisted cytologic interpretation combined with HPV genotyping in the pathological diagnosis of cervical lesions

SHANG Haixia; SHI Xiaofeng; YU Hongxin; FENG Yaqian; HUANG Yue; GUO Qi; WANG Xiaoxue; HU Lina; WANG Lin

doi:10.3969/j.issn.1007-8096.2025.12.006

Abstract：Objective To evaluate the diagnostic performance and clinical application value of artificial intelligence(AI)-assisted cytologic interpretation combined with human papillomavirus(HPV)genotyping in the diagnosis of cervical intraepithelial neoplasia grade 2(CIN2+)or worse,thereby informing optimization of screening strategies.Methods A retrospective paired diagnostic efficacy study was conducted,including 300 cervical liquid-based cytology(LBC)specimens and corresponding histopathological results from women who attended Shanxi Bethune Hospital between March 2021 and March 2025.The diagnostic sensitivity,specificity,positive predictive value(PPV),negative predictive value(NPV),and area under the receiver operating characteristic curve(AUC)in the diagnosis of CIN2+were compared among four approaches:the AI cytology model,pathologist manual review,HPV16/18 genotyping,and the AI+HPV combined model.Differences were compared using the DeLong and McNemar tests.Clinical net benefit across risk thresholds was assessed using decision curve analysis(DCA).Results For CIN2+diagnosis,AI cytology model yielded a sensitivity of 87.2%,specificity of 80.4%,and AUC of 0.90(95%CI:0.86-0.94).Pathologist manual review achieved a sensitivity of 85.0%,specificity of 88.9%,and AUC of 0.91(95%CI:0.87-0.95).HPV16/18 genotyping showed the highest sensitivity(90.2%)but the lowest specificity(56.5%),with an AUC of 0.80(95%CI:0.75-0.84).The AI+HPV combined model demonstrated the best performance,with a sensitivity of 95.0%,specificity of 86.2%,PPV of 90.0%,NPV of 96.2%,and AUC of 0.95(95%CI:0.92-0.97),significantly outperforming any single method(all P<0.01).DCA indicated that the combined model consistently provided the greatest net benefit within the 10%-40%threshold probability range.Subgroup analysis confirmed that the combined model maintained stable performance across different HPV genotypes,age groups,and cytology categories.Notably,the combined model markedly improved NPV(96.2%)in ASC-US/LSIL cases and achieved a sensitivity of 98.0%in HSIL and above cases.Conclusion The combined diagnostic model of AI-assisted cytologic interpretation and HPV genotyping achieves both high sensitivity and specificity,reducing missed diagnoses while avoiding excessive referrals.It offers greater clinical net benefit than single-method approaches and shows promise as a multimodal strategy in cervical cancer screening and pathological diagnosis.