A Hybrid CNN–Transformer Framework for Accurate Cervical Cancer Diagnosis Using Multiscale Cytology Features

Cervical cancer is one of the most preventable yet widely prevalent cancers among women, particularly in low- and middle-income regions where access to screening remains limited. Although preventive measures such as human papillomavirus (HPV) vaccination and Pap smear tests are available, many cases are still detected at advanced stages, increasing mortality rates. Conventional diagnosis relies on manual examination of cytology slides, which is time-consuming, subjective, and prone to variability, highlighting the need for automated and reliable systems. Deep learning has shown strong potential in medical image analysis. Convolutional Neural Networks (CNNs) effec-tively capture local features such as nuclear morphology and texture but struggle with global context due to limited receptive fields. On the other hand, Transformer-based models excel at capturing long-range dependencies through self-attention but lack the ability to focus on fine-grained local details. To overcome these limitations, this study proposes a hybrid CNN–Transformer framework that integrates local and global feature extraction using a multiscale fusion strategy. Evaluated on the Herlev dataset, the model achieves improved performance across key metrics. Additionally, Grad-CAM is used for interpretability, making the system more reliable for clinical applications and early cervical cancer detection.