Enhanced Diagnosis of Cervical Cancer Using Archer Fish Hunting Optimizer-Optimized Xception Network with AI Interpretability via Grad-CAM.

OBJECTIVE: The primary objective of this study is to enhance the performance of cervical cancer classification by optimizing the Xception deep learning model using the Archer Fish Hunting Optimizer (AHO). The study aims to evaluate the effectiveness of the AHO-optimized Xception model in classifying cervical cytology images and to compare it against traditional CNN architectures. It also integrates Grad-CAM visualization to ensure model interpretability and to support explainable AI in clinical decision-making. METHODS: A cervical cytology image dataset comprising four diagnostic classes High Squamous Intraepithelial Lesion (HSIL), Low Squamous Intraepithelial Lesion (LSIL), Negative for Intraepithelial Malignancy (NILM), and Squamous Cell Carcinoma (SCC) was used for model development. Preprocessing steps included image normalization and augmentation to improve generalization. Several deep learning models AlexNet, MobileNet V2, ResNet-50, Inception V3, and the standard Xception were trained and benchmarked. The AHO algorithm was applied to optimize the hyperparameters of the Xception model. Model performance was evaluated using metrics such as accuracy, precision, and AUC-ROC. Grad-CAM visualization was utilized to highlight image regions influencing classification decisions. RESULT: The AHO-optimized Xception model demonstrated superior classification performance, achieving 98.96% accuracy, 98.57% precision, and an AUC-ROC of 99.97%. It outperformed all baseline models in classification effectiveness. The Grad-CAM-based visualization provided valuable insights into the model's focus areas, supporting interpretability and confidence in its predictions. CONCLUSION: The study establishes the efficacy of using the Archer Fish Hunting Optimizer for hyperparameter tuning in deep learning-based cervical cancer diagnosis. The AHO-optimized Xception model offers improved classification performance and robust explainability, making it a valuable tool for AI-assisted cervical cancer screening. Future research may explore hybrid metaheuristic strategies and their real-time deployment in clinical environments.