Explainable automated classification of cervical cancer cells using SE-LGANet with squeeze and excitation blocks and local and global attention mechanisms
TL;DR
A Convolutional Neural Network integrated with Squeeze-and-Excitation blocks and dual attention mechanisms was developed to classify five distinct cell types, providing an explainable and systematically validated approach for cervical cell evaluation.
Abstract
Cervical cancer remains a significant global health threat, making early detection through Pap Smear screenings essential for improving survival rates. This study aims to enhance the classification of cervical cell categories using the SIPaKMeD dataset, focusing on increasing diagnostic accuracy and model interpretability. A Convolutional Neural Network (CNN) integrated with Squeeze-and-Excitation (SE) blocks and dual attention mechanisms was developed to classify five distinct cell types: Koilocytotic, Parabasal, Superficial-Intermediate, Dyskeratotic, and Metaplastic cells. Grad-CAM was utilized to visualize lesion areas and ensure model interpretability. The architecture was further validated using Mendeley LBC and Herlev datasets. The proposed model achieved 91.48% accuracy, 91.57% sensitivity, and 97.87% specificity on the SIPaKMeD dataset. It also demonstrated high transferability with 96.88% accuracy on Mendeley LBC and 89.62% on the Herlev dataset. Grad-CAM, LIME, and SHAP confirmed the model’s focus on clinically relevant features. The integration of attention mechanisms and interpretability tools provides an explainable and systematically validated approach for cervical cell evaluation. This framework bridges the gap between AI performance and clinical trust, with future research focusing on real-world clinical integration.
