Attention-Enhanced DenseNet Architecture for Robust Cervical Cytology Classification

Abstract

This study presents an attention-enhanced DenseNet201 framework for cervical cytology classification. Four attention mechanisms, SAM, CA, CBAM, and BAM, were integrated into the DenseNet201 backbone and evaluated with four optimizers (Adam, AdamW, RMSProp, and SGD). Experiments were conducted on the Mendeley and Pomeranian multicell dataset using percentage split and 5 fold cross-validation. In multi-class classification, CA with Adam achieved the highest accuracy of 98.75%, while several attention optimizer combinations reached 99.90% in binary classification. On the Pomeranian dataset, crossvalidation accuracy reached 90.04%, confirming the model’s generalization capability across datasets of different sizes and class distributions. These results demonstrate that attention-enhanced DenseNet architectures effectively emphasize diagnostically relevant regions, providing a reliable and robust approach for automated cervical cytology analysis.