Enhancing Cervical Cancer Detection: Explainable AI and Attention Mechanisms for Pap Smear Classification

Cervical cancer is the second most prevalent female cancer in India, responsible for almost $\mathbf{1 8. 3 \%}$ of all cancers in women, with more than 120,000 new cases every year. Though preventable and curable at early ages, poor availability of skilled cytopathologists and late diagnosis remain the causative factors of high mortality. Conventional approaches to diagnosis heavily depend on Pap smear slide manual screening, which is time consuming, subjective, and prone to errors, particularly in resourcescarce areas. To address these challenges, this study proposes an attention-based and explainable deep learning architecture for cervical cell classification. The architecture is constructed over the EfficientNet-B0 backbone for strong feature extraction, supplemented with a Convolutional Block Attention Module (CBAM) to highlight clinically important features by using improved channel-wise attention. The system has an extensive preprocessing pipeline including image resizing, denoising, Otsu’s thresholding, morphological operations, and color normalization to provide quality input data. The model is trained on a curated dataset comprising both cancerous and normal cervical cytology images. Post-hoc explanation methods such as Grad-CAM++, SHAP, and LIME are used for explaining and justifying the predictions made by the model. The proposed method attains an accuracy of 82%, a sensitivity of 80.6%, specificity of 83.1%, and an AUC score of $\mathbf{0. 8 7}$. This work emphasizes the potential of explainable deep learning systems to aid early and accurate cervical cancer diagnosis, particularly in underserved healthcare environments.