Automated and hybrid artificial intelligence cervical imaging workflows demonstrated high specificity and negative predictive value but exhibited limited sensitivity when evaluated against routine visual inspection with acetic acid in community-based screening. The multicenter investigation examined the concordance of algorithmic image evaluation systems with routine visual screening across multiple healthcare facilities, finding that both technological approaches demonstrated notable limitations in identifying positive cases identified by clinical staff.
Visual inspection with acetic acid remains widely utilized for cervical precancer screening throughout low- and middle-income countries, serving as an accessible frontline method. However, the procedure is recognized to be subject to operator-dependent variability between different healthcare providers. Although automated cervical imaging systems have been introduced with the goal of improving diagnostic consistency, their performance across heterogeneous real-world clinical settings has remained insufficiently characterized.
To evaluate these technologies under field conditions, researchers conducted a multicenter cross-sectional study that enrolled 253 women undergoing cervical screening across four facilities in West Java, Indonesia. The screening evaluations took place between October 2025 and January 2026. Within the study framework, routine visual inspection with acetic acid was employed as the operational comparator rather than an established histopathological reference standard.
The study protocol assessed cervical images captured both before and after the application of acetic acid. These paired images were processed through two separate screening workflows: an automated artificial intelligence system and a hybrid workflow that incorporated expert clinical review. Screening results from both workflows were dichotomized into positive or negative classifications, and concordance relative to routine visual inspection was evaluated using Cohen’s kappa, sensitivity, specificity, positive predictive value, negative predictive value, and overall agreement.
Across the cohort of 253 women, routine visual inspection identified 23 participants, or 9.1 percent of the total sample, as positive. Compared against these routine inspection results, the automated artificial intelligence system demonstrated an overall agreement of 88.9 percent and moderate concordance, with a Cohen’s kappa of 0.421 and a 95 percent confidence interval of 0.266 to 0.622. The automated system achieved a sensitivity of 56.5 percent with a 95 percent confidence interval of 34.5 to 76.8 percent, a specificity of 92.2 percent with a 95 percent confidence interval of 87.9 to 95.3 percent, a positive predictive value of 41.9 percent, and a negative predictive value of 95.5 percent.
The hybrid workflow incorporating expert review likewise reached an overall agreement of 88.9 percent, but showed only fair concordance with a Cohen’s kappa of 0.206 and a 95 percent confidence interval of 0.014 to 0.398. Relative to visual inspection, the hybrid workflow showed a sensitivity of 21.7 percent with a 95 percent confidence interval of 7.5 to 43.7 percent, a specificity of 95.7 percent with a 95 percent confidence interval of 92.1 to 97.9 percent, a positive predictive value of 33.3 percent, and a negative predictive value of 92.4 percent. While site-specific agreement varied across the four facilities, analyses at individual locations were limited by the sparse number of positive cases.
The researchers concluded that both the automated and hybrid workflows provided high specificity and high negative predictive value but suffered from limited sensitivity when evaluated against routine visual inspection with acetic acid. The findings emphasize the performance variability of artificial intelligence screening across real-world environments and indicate the critical role of standardized image acquisition. The authors noted that histopathological validation remains necessary to establish true diagnostic accuracy and clinical utility.
Subject of Research: Concordance of AI-assisted cervical imaging systems with visual inspection with acetic acid
Article Title: Multicenter evaluation of concordance between AI-assisted and hybrid AI-assisted cervical imaging systems and visual inspection with acetic acid in community-based cervical precancer screening
Article References: Zenyda, R. R. S., Alhayat, M. R. V. A., Wardhana, R. M. W., Adzkia, Q., Nabiylah, N. A., Zahran, M. G. S., Mukhayadi, H., Suardi, D., Mantilidewi, K. I., Agustian, D., & Hidayat, Y. M. (2026). Multicenter evaluation of concordance between AI-assisted and hybrid AI-assisted cervical imaging systems and visual inspection with acetic acid in community-based cervical precancer screening. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-03002-x
Image Credits: AI Generated
DOI: 10.1186/s12880-026-03002-x
Keywords: cervical precancer screening, visual inspection with acetic acid, AI-assisted cervical imaging, hybrid AI workflow, Cohen's kappa, diagnostic concordance, West Java, Indonesia, cross-sectional study, screening sensitivity, screening specificity, positive predictive value, negative predictive value, operator variability, standardized image acquisition, histopathological validation, multicenter evaluation, community-based screening
News Source: Ophelia Keating. (November 14, 2026). AI Cervical Screening Shows High Specificity but Limited Sensitivity in Indonesian Study. Scienmag.



