Accelerated and accurate cervical cancer diagnosis using a novel stacking ensemble method with explainable AI

Cervical cancer is a preventable yet life-threatening disease that claims hundreds of thousands of lives each year, particularly in low-resource settings where timely screening is scarce. Current Deep Learning (DL) approaches for automated cervical cytology classification encounter challenges such as class imbalance, computational inefficiency, and inadequate generalizability. This study proposes a novel CerviXEnsemble model that integrates multiple pre-trained DL architectures (Inception-ResNetV2, EfficientNet-B6, ResNet152, Inception-ResNetV2, EfficientNet-B6, DenseNet201, and NASNetMobile) as base learners, along with a dense-layer meta-learner that refines and consolidates predictions for improved robustness. Unlike traditional single-CNN models, our stacking ensemble approach utilizes diverse feature representations to enhance classification stability and generalization across multiple cytology datasets. To validate the model, we experimented with the Herlev and SIPaKMeD benchmark datasets in this study. Techniques like contrast enhancement and data augmentation were employed to optimize feature extraction. The model achieved state-of-the-art performance, attaining an accuracy of 99.38 % and an F1-score of 98.49 % on the Herlev dataset and an accuracy of 98.71 % and an F1-score of 97.53 % on SIPaKMeD. These performances are superior to previous studies in controlling class imbalance and providing stable predictions over different samples. Additionally, Explainable AI (XAI) techniques were incorporated to ensure transparent and interpretable predictions, aiding clinicians in their decision-making processes. An interpratable web application was developed for real-time Pap smear analysis to reduce the diagnostic workload for pathologists by identifying high-risk samples. This solution shows great promise for use in various healthcare settings, maintaining high diagnostic accuracy while requiring minimal computational resources, making it suitable for both urban hospitals and rural clinics.