ACHS-CervNet: A Hybrid Stacked Deep Learning Approach for Accurate Cervical Cancer Detection and Classification

Abstract

Cervical cancer remains one of the leading causes of mortality among women worldwide, and early, accurate detection is crucial for effective treatment. Conventional cytological screening suffers from inter-observer variability and limited accuracy, prompting increased adoption of artificial intelligence, Machine Learning (ML), and Deep Learning (DL) methods. However, existing approaches face challenges, including insufficient feature discrimination, inter-class similarity, and limited generalization across datasets. To address these gaps, this study formulates the problem of designing a robust and generalizable model for cervical cancer detection and classification. The primary objectives of this work were to develop a Convolutional Neural Network (CNN)-based correlation feature map for accurate feature representation and to design a hybrid stacked DL framework for multi-class cervical cancer classification. Hence, this work presents the Attention-Correlation Hybrid Stacked Cervical Network (ACHS-CervNet), which integrates per-channel-spatial attention with stacked ResNet50 and DenseNet121, enhancing feature selection and classification performance. Experimental evaluations on the SIPaKMeD and Herlev datasets demonstrated the effectiveness of ACHS-CervNet. For SIPaKMeD, the ACHS-CervNet achieved 99.82% accuracy on 2-class, 99.85% on 3-class, and 99.92% on 5-class classification tasks. Similarly, on the Herlev dataset, it achieved 99.8% accuracy for 2-class and 99.65% for 7-class classification, outperforming existing state-of-the-art models.