Abstract
BACKGROUND: Cervical cancer remains a major cause of morbidity and mortality among women, particularly in low- and middle-income countries (LMICs) where screening access is limited. Visual Inspection with Acetic Acid (VIA) is widely used but subject to inter-observer variability. Artificial intelligence (AI)-assisted cervical imaging may improve standardization; however, real-world community evidence is limited. OBJECTIVE: To evaluate the concordance of AI-assisted and hybrid AI-assisted cervical imaging systems (CIS) relative to VIA findings in a community-based cervical precancer screening program in West Bandung, Indonesia. METHODS: A cross-sectional study was conducted among 71 women to assess concordance and operational performance of two AI-based approaches: an AI-assisted CIS (Cerviray AI(®), AIDOT Inc., Seoul, South Korea) and a hybrid AI-assisted CIS incorporating expert review within a VIA-based workflow. RESULTS: The AI-assisted CIS showed sensitivity of 66.7% (95% CI 20.8-93.9) and specificity of 95.6% (95% CI 87.8-98.5), with PPV 40.0% and NPV 98.5%. The hybrid approach demonstrated similar sensitivity (66.7%; 95% CI 20.8-93.9) and slightly higher specificity (97.1%; 95% CI 89.9-99.2), with PPV 50.0% and NPV 98.5%. VIA positivity was low (4.2%; 8.5% using a composite definition), likely inflating specificity and NPV. All metrics were calculated against VIA rather than histopathology. Agreement with VIA was moderate (κ = 0.472 and κ = 0.550). CONCLUSION: AI-assisted and hybrid CIS showed moderate agreement with VIA. Interpretation is limited by small sample size, wide confidence intervals for sensitivity, low disease prevalence, and use of VIA as the reference standard. These preliminary findings suggest a potential supportive role for AI-assisted imaging in LMIC screening, warranting validation in larger studies using histopathologic reference standards.