Artificial intelligence use in the routine of cervical-vaginal cytology: a systematic review

Abstract Cervical cancer remains a major global public health problem, although it is largely preventable through vaccination against human papillomavirus and organized screening. Cervical cytology has historically been the primary method of screening, but this method has limitations related to its observer-dependent nature, intra- and inter-observer variability, and the diagnostic service burden. In this context, artificial intelligence has emerged as a promising tool to automate and standardize cytological analysis. The objective of this systematic review was to critically synthesize the available evidence on the application of artificial intelligence in cervical cytology, focusing on diagnostic performance, methodological limitations, and implications for screening programs. Search was conducted on the PubMed/MEDLINE base, resulting in the inclusion of 30 published studies in the last five years. Publications have predominantly analyzed deep learning algorithms, especially convolutional neural networks, applied to cytological classification according to the Bethesda System. In general, studies have shown high sensitivity and accuracy of AI-assisted cytology in the detection of squamous intraepithelial lesions and invasive carcinoma, often with comparable or superior performance to conventional cytological reading. However, specificity showed wide variation across studies, reflecting methodological heterogeneity, differences in datasets, and limitations related to external validation. Key constraints identified included the predominance of retrospective designs, the use of partial gold patterns, and challenges associated with scanning cytological samples. In summary, the application of artificial intelligence to cervical cytology shows significant potential to enhance the sensitivity, standardization, and efficiency of cervical cancer screening programs. However, its broad incorporation into clinical practice should occur in integration with human expertise and depends on the conduct of prospective, multi-center studies, with robust external validation and impact assessment in real-world scenarios.