Performance evaluation of artificial intelligence–assisted diagnostic tools for human papillomavirus–related cervical and anal cancers and their precancerous lesions: A systematic review and meta-analysis

Detection of high-grade squamous intraepithelial lesions (HSILs) is key for the prevention of human papillomavirus (HPV)–related cancers. In this study, we aimed to identify and consolidate the existing evidence on the diagnostic performances of artificial intelligence (AI)–assisted tools for HPV-related HSILs. We followed the Preferred Reporting Items for a Systematic Review and Meta-analysis of Diagnostic Test Accuracy guidelines and systematically searched PubMed, Embase, Web of Science, IEEE, and the Cochrane Library for eligible studies published between 2000 and January 2024. Studies on AI-assisted colposcopy used to facilitate the diagnosis of cervical or anal HSILs and cancers were included. The revised Quality Assessment of Diagnostic Studies 2 and AI checklists were used to assess the risk of bias. With random-effect models, meta-analyses were performed on performance indicators, including accuracy, sensitivity, specificity, and area under the receiver-operating characteristic (AUC) curves. Subgroup and meta-regression analyses were performed to identify the sources of heterogeneity. Twenty-five studies, comprising 21 studies focused on cervical cancer and 4 on anal cancer, were included in the meta-analysis. When differentiating from individuals with <CIN2/LSIL− (lesion grade lower than cervical intraepithelial neoplasia grade 2 or lower than or equal to low-grade squamous intraepithelial lesion), the AI-assisted diagnostic tools identified individuals with CIN2+/HSIL+ (lesion grade ≥ CIN2 or ≥ high-grade squamous intraepithelial lesion) at a pooled accuracy of 0.84 (95% confidence interval [CI]: 0.81–0.88), sensitivity of 0.87 (0.81–0.93), specificity of 0.85 (0.81–0.89), and AUC of 0.94 (0.91–0.96). For anal cancer, the AI-assisted diagnostic tools differentiated individuals with AIN2+/HSIL+ (≥AIN2 or ≥HSIL) from those with <AIN2/LSIL− (lesion grade < anal intraepithelial neoplasia grade 2 or ≤ LSIL) at a pooled accuracy of 0.90 (95% CI: 0.88–0.93), sensitivity of 0.90 (0.85–0.94), specificity of 0.88 (0.71–0.98), and AUC of 0.97 (0.95–0.98). All meta-analysis results demonstrated high heterogeneity ( I 2 > 90%), but the meta-regression analysis did not reveal any significant contributing sources of heterogeneity. Based on this systematic review of existing literature, we concluded that AI-assisted diagnostic tools have demonstrated promising predictive performance in diagnosing HPV-related cervical and anal precancerous lesions and cancers.