A Hybrid Deep Learning Framework for Cervical Cancer Classification Using Multi-Scale Feature Fusion and Autoencoder Compression

Abstract

Amidst the rapid growth of research and technology in cervical cancer detection, ensuring accuracy, especially with varied sample quality in traditional Pap smears, is still one of the major challenges. Automated classification of cervical cell images facilitates early detection of abnormal cellular changes critical for cervical cancer screening. This paper proposes a hybrid, multi-class cervical cell classification that integrates segmentation, deep feature extraction, feature compression, and ensemble classification. Initially, segmentation was done by training a Swin-UNETR using pseudo masks. These segmented images served as an input for the two main CNNs used for deep feature extractions at multiple resolutions. The fused features were subsequently refined and compressed using an MLP autoencoder. Classification was then performed by a weighted ensemble model incorporating both a Support Vector Machine and a Random Forest. Experiments were carried out on the SIPaKMeD dataset and achieved a test accuracy of 95.55%, confirming robust performance across all five cervical cell types. Therefore, based on the results, we can infer that combining segmentation, deep feature extraction at multiple resolutions, autoencoder’s latent features and ensemble learning can be a practical and effective approach for automated cervical cell analysis.