Swin Transformer–Driven Cervical Cell Classification with Explainable AI and Web-Based Screening

Abstract

Accurate interpretation of cervical cytology images is essential for effective cervical cancer screening, yet manual assessment is time-consuming and subject to observer variability. This paper presents a transformer-based deep learning framework for automated cervical cell classification using Pap smear images. We conduct a systematic evaluation of modern attention-driven architectures, including MaxViT, Swin Transformer, EfficientFormer, and HorNet, under a unified preprocessing and training pipeline designed to handle staining variability and class imbalance. To enhance model transparency and clinical trust, explainable AI is integrated via Grad-CAM, enabling visual localization of cytomorphological regions that drive model decisions. Experiments on the Herlev and SIPaKMeD datasets demonstrate that the proposed Swin Transformer achieves superior and consistent performance, reaching 99.27% accuracy on Herlev and 98.82% accuracy on SIPaKMeD, with high MCC and PR-AUC values. In addition, a lightweight web-based application is developed to support dataset selection, real-time inference, confidence reporting, and visual explanation. The results confirm that hierarchical transformer architectures can deliver accurate, interpretable, and deployable solutions for computer-aided cervical cancer screening.