CITOBOT AI for real-world cervical cancer screening using colposcopy imaging

Cervical cancer remains a major cause of morbidity and mortality among women, particularly in low- and middle-income countries, where delays in screening and diagnostic follow-up limit early detection. Artificial intelligence applied to medical imaging may strengthen screening programs by improving accuracy, consistency, and timeliness in resource-limited settings.

This study aimed to evaluate the internally validated screening performance of CITOBOT AI, an artificial intelligence system for real-world cervical cancer screening using colposcopy imaging, in a public hospital setting in Colombia.

A cross-sectional study was conducted among 650 women screened at 'Siloé' Hospital in Cali, Colombia, between February 2023 and July 2025. Colposcopy-guided biopsy served as the reference standard. A total of 2,648 cervical images were classified using a predefined binary endpoint: screen-negative/no risk versus screen-positive/at risk. The model was developed using transfer learning, image segmentation, data augmentation, patient-level data partitioning, and five-fold cross-validation within the training subset. Screening performance was evaluated using accuracy, sensitivity, specificity, area under the receiver operating characteristic curve, predictive values, and confusion matrix analysis. A secondary exploratory analysis examined associations between selected clinical variables and AI-based screening classification.

CITOBOT AI achieved internally validated screening performance with an accuracy of 94.3%, sensitivity of 93.4%, specificity of 94.9%, and an area under the receiver operating characteristic curve of 0.98. Positive and negative predictive values were 92.9 and 96.2%, respectively. Performance estimates were stable across patient-level folds and consistent with the hold-out validation subset within the same dataset. HPV status showed an unexpected inverse association with AI screen-positive classification; this finding should not be interpreted as biologically protective or as evidence that HPV status influenced model output, since HPV status was not used as an input variable.

CITOBOT AI demonstrated promising internally validated performance for real-world cervical cancer screening based on colposcopy imaging. The predefined binary classification is consistent with its intended role as a screening support tool to identify women who may require confirmatory colposcopy and biopsy, while histopathology remains the reference standard for lesion grading and therapeutic decision-making. External validation in independent multicenter and population-based datasets is required before broader implementation.