An Explainable Multi-Level Framework for Cervical Cancer Detection Using Traditional Computer Vision and Deep Learning

Cervical cancer is one of the most common causes of mortality among women globally. Development of accurate, interpretable and clinically deployable automated systems for detection of precancerous cells is essential. We propose an explainable multi-level framework that integrates traditional computer vision techniques, ML models, and DL models. We use three benchmark datasets for evaluating our framework, namely SIPaKMeD, Mendeley LBC, and Herlev. The traditional computer vision techniques used are Histogram of Oriented Gradients (HOG), Local Binary Patterns (LBP), and Scale-Invariant Feature Transform with Bag-of-Visual-Words (SIFT-BoVW). These features are combined with ML models that include Logistic Regression, Random Forest, and Linear SVM.We also use three pretrained models, namely AlexNet, ResNet50, and EfficientNet-B0, along with a custom CNN–Transformer. Each model was evaluated across five cross-validation folds.Within theMLbranch, the Random Forest with LBP features achieved the highest mean accuracies on all the three datasets, with 72.27% on SIPaKMeD, 80.77% on Mendeley LBC, and 47.98% on the Herlev dataset. Within the DL branch, ResNet50 had the best performance among all the models, achieving mean accuracies of 97.55% on SIPaKMeD, 99.48% on Mendeley LBC, and 71.53% on Herlev. Explainable AI techniques such as Feature Importance (for ML models) and Grad-CAM and Integrated Gradients (for DL models) were incorporated to ensure interpretability across all models. The XAI analysis was structured across four dimensions: models, features, datasets, and lesion grades to evaluate the effectiveness of the framework at multiple levels.