Application of Artificial Intelligence in Cervical Cytology: A Systematic Review of Deep Learning Models, Datasets, and Reported Metrics

Miguel Angel Valles-Coral^1*Ciro Rodriguez^2*Diego Rodriguez³Keller Sánchez-Dávila⁴Lolita Arévalo-Fasanando⁵Nelly Reátegui-Lozano⁵

• ¹Faculad de Ingeniería de Sistemas e Informática, National University of San Martan, Tarapoto, Peru
• ²Facultad de Ingeniería de Sistemas e Informática, Universidad Nacional Mayor de San Marcos, Lima District, Peru
• ³Facultad de Ciencias de la Salud, Medicina, Universidad Peruana de Ciencias Aplicadas, Lima District, Peru
• ⁴Faculad de Medicina Humana, National University of San Martan, Tarapoto, Peru
• ⁵Faculad de Ciencias de la Salud, National University of San Martan, Tarapoto, Peru

The use of artificial intelligence (AI) in cervical cytology has grown significantly, driven by the need to automate the early diagnosis of precancerous lesions. This systematic review analyzes recent studies on deep learning models applied to cytological images, focusing on the architectures used, the datasets employed, and the performance metrics reported. The PRISMA methodology was used to select articles published between 2022 and 2025 in Scopus. Following a rigorous selection and analysis process, 77 studies were included for RQ1 (models), 75 for RQ2 (datasets), and 71 for RQ3 (metrics), all of which met the eligibility criteria. The results show a predominance of hybrid models (56%), followed by pure convolutional neural networks (CNNs), and a growing adoption of Vision Transformers (ViTs). The most frequently used datasets were SIPaKMeD and Herlev, although there is an increasing trend toward the use of private or proprietary datasets. The most commonly reported metric was accuracy, with an average of 87.76%, followed by precision, recall, and F1-score. Hybrid and ViT-based models demonstrated the best performance, exceeding 92% accuracy in several cases. However, frequent limitations were identified, such as limited cross-validation, the use of images that poorly reflect real clinical settings, and a lack of standardization in diagnostic criteria. This review synthesizes the current landscape and proposes guidelines for future research aimed at integrating artificial intelligence (AI) models into real clinical environments.