Abstract
Manual review of cervical cytology specimens is labor-intensive and susceptible to fatigue-related errors, which has motivated the development of computer-aided diagnosis (CAD) systems for cervical cancer screening. To address the limited ability of conventional convolutional neural networks (CNNs) to capture global contextual relationships in cellular morphology, this study comparatively evaluates four contemporary Vision Transformer architectures, ViT-Base, DeiT-Base, PiT-Base, and Swin-Base, for five-class cervical cell classification. Experiments were conducted on the SIPaKMeD dataset comprising 4,049 isolated cell images. ImageNet-based transfer learning and strong regularization through RandAugment, Mixup, and CutMix were used to mitigate overfitting under limited-data conditions. Performance was assessed using accuracy, precision, recall, macro F1-score, parameter count, and computational complexity. ViT-Base achieved the best overall classification performance, with an accuracy of 96.90\% and a macro F1-score of 0.9691. PiT-Base provided the most favorable computational profile, requiring 72.74 million parameters and 21.1067 GFLOPs while maintaining an accuracy of 96.57\%. Error analysis showed that most residual confusion occurred between the morphologically similar metaplastic and koilocytotic classes. These findings indicate that Vision Transformer models can provide accurate and reproducible decision support for cervical cytology; however, external validation on multicenter and whole-slide data remains necessary before clinical deployment.