An amalgamation of deep neural networks optimized with Salp swarm algorithm for cervical cancer detection

• Published in: Computers & electrical engineering Vol. 123; p. 110106
• Main Authors: Bilal, Omair, Asif, Sohaib, Zhao, Ming, Khan, Saif Ur Rehman, Li, Yangfan
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.04.2025
• Subjects: Cancer cells classification; Cervical cancer disease; Deep learning; Ensemble learning; Salp swarm algorithm; Deep learning; Salp swarm algorithm; Cervical cancer disease; Ensemble learning; Cancer cells classification
• ISSN: 0045-7906
• DOI: 10.1016/j.compeleceng.2025.110106

Cervical cancer is a critical global health issue by affecting millions of women each year and causing high mortality rates if not diagnosed early. Early detection of cervical cancer significantly improves patient outcomes and survival rates. Traditional diagnostic approaches are frequently susceptible to errors, which can result in inaccurate diagnoses and being time-consuming. However, the emergence of machine learning and artificial intelligence provides innovative approaches to improve both diagnostic accuracy and efficiency. Individual deep learning models often face challenges in extracting the critical information essential for accurate prediction of disease in complex datasets. To tackle this, we propose a novel ensemble model that leverages the Salp Swarm Algorithm (SSA) to enhance cervical cancer diagnosis. This approach employs three highly effective pre-trained models like DenseNet169, DenseNet201, and Xception for feature extraction. To improve feature attention, we integrated the Convolutional Block Attention Module into each of these models to make them our base models. Subsequently, the predictions generated from each base model within the ensemble are aggregated through a weighted aggregation approach and further optimized the ensemble model by intelligently assigning weights to each model through the SSA. We evaluated our model using two datasets including the Mendeley LBC dataset with four classes and the BloodMNIST Benchmark dataset with eight classes. This approach ensures the robustness and generalizability of the ensemble model by demonstrating its effectiveness on diverse datasets. Our proposed ensemble model demonstrates superior performance compared to existing state-of-the-art methods by attaining an impressive accuracy rates of 99.48% on the 4-class Mendeley LBC dataset and 95.23% on the 8-class BloodMNIST dataset. This work marks a significant advancement in the field of cervical cancer diagnosis. We evaluate our optimized ensemble model using advanced metrics and visualizations such as confusion matrix, receiver operating characteristics (ROC) curve, t-distributed Stochastic Neighbor Embedding (t-SNE) plot, and Grad-CAMs. We validated the significance of our findings by conducting McNemar's Chi-Square and Friedman's Test. This comprehensive assessment underscores the accuracy, robustness, and interpretability of the proposed model in diagnosing cervical cancer cells.