Cervical cancer remains one of the leading causes of mortality among women worldwide despite being largely preventable through early detection and vaccination. Conventional cytology screening requires skilled personnel and is prone to subjectivity, motivating the integration of artificial intelligence tools for automated cytological analysis. This paper presents a comprehensive evaluation of four convolutional neural network (CNN) models: InceptionV3, MobileNetV2, VGG16, and DenseNet121, for automated classification of cervical cells into Normal and Suspect categories. Experiments were conducted using the public SIPaKMeD dataset and a newly collected proprietary clinical cytology dataset. The models were trained using a two-phase training and fine-tuning approach to improve generalization across domains, and performance was statistically evaluated using confidence intervals. On the SIPaKMeD test set, InceptionV3 and DenseNet121 achieved 98-99% accuracy, while MobileNetV2 and VGG16 reached 96-83%. On the proprietary dataset, VGG16 and InceptionV3 maintained 89% and 86% accuracy, respectively, confirming their robustness despite increased real-world variability. These findings highlight the domain shift between public and proprietary cytology data and demonstrate that selected CNN architectures can effectively support cytological screening as decision-support tools in clinical environments.