Quantile Deviation Ensemble Based on Multi-Layer Perceptrons for Cervical Cancer Classification with Uncertainty Perception

The accurate classification of cervical cytology in Pap smear images remains a critical challenge in computer-aided diagnosis, largely due to the inherent uncertainty and subtle morphological variations among different pathological categories. To address this, we propose a novel uncertainty-aware ensemble framework that integrates statistical quantile analysis with deep learning for robust and interpretable classification. Our framework first leverages three deep convolutional neural networks DenseNet121, MobileNetV2, and ResNet-50 as base feature extractors. Instead of employing naive ensemble strategies, we introduce a quantile deviation based weighting mechanism to dynamically assess and integrate the prediction confidence of each model, explicitly quantifying performance bias across different probability quantiles. This approach not only enhances ensemble stability but also provides a statistical measure of model uncertainty. Subsequently, the weighted probabilistic outputs are fed into a multi-layer perceptron (MLP) for further non-linear optimization and decision refinement, forming a hybrid statistical deep learning pipeline. Evaluated on the publicly available SIPaKMeD dataset, our framework achieves an average accuracy of 98.10%, outperforming both individual base models and existing ensemble methods. Visualization via Grad-CAM further confirms that the framework focuses on clinically relevant cellular structures, validating its diagnostic relevance. By bridging statistical uncertainty quantification with deep ensemble learning, this work offers a principled and transparent methodology for medical image classification, with potential extensibility to other domains requiring reliable and interpretable predictions under uncertainty.