Hybrid DL Classification Model Based on CNN and Transformer for Pap Smear Cells

Cervical cancer is a leading cause of cancer-related deaths among women, with early detection via Pap smear screening significantly reducing mortality. However, traditional analysis is resource-intensive and expertise-dependent, posing challenges in low-resource settings. The Pap Smear Cell Classification Challenge (PS3C) at of the International Symposium on Biomedical Imaging (ISBI) 2025 aims to advance automated cervical cell classification into three categories: Healthy, Rubbish, and Unhealthy. We propose a hybrid deep learning model integrating CNN and Transformer architectures to leverage Convolutional Neural Network's (CNN) local feature extraction and the Transformer's global context modeling. Two variants-CNN-vision Transformer (CNN-ViT) and CNN-Hierarchical Vision Transformer using Shifted Windows (CNN-Swin)-were developed, with CNN-Swin outperforming CNN-ViT, achieving an F1 score of 0.83175 in the first test phase. The Unhealthy class had lower classification accuracy due to class imbalance, which we mitigated using weighted categorical focal loss. Further, an ensemble of top CNN-Swin models improved robustness, securing third place in the final test phase rankings with an F1 score of 0.79446. To enhance performance, we recommend expanding the Unhealthy class dataset or generating synthetic samples using Generative Adversarial Networks.