Explainability-Aligned Reliability-Weighted Fuzzy Ensemble for Automated Cervical Cancer Classification

Abstract

Cervical cancer remains a major global health concern, highlighting the need for computer-aided diagnostic systems that are both reliable and interpretable. Despite advances in deep learning–based cytology image classification, a gap persists in aligning model predictions with biologically meaningful explanations. This study aims to develop an explainability-aligned, sample-wise reliability-weighted fuzzy ensemble framework for cervical cytology image classification to enhance both performance and interpretability.

Methods

The proposed framework integrates three pretrained convolutional neural network backbones—InceptionV3, MobileNetV2, and Inception-ResNetV2—within a fuzzy ensemble structure. A novel explainability metric, termed Explainable Artificial Intelligence Alignment (XAIHit), is introduced to quantitatively assess the spatial correspondence between Grad-CAM activation maps and annotated cytoplasmic and nuclear regions. The model combines calibrated confidence estimates with XAIHit to produce a per-sample reliability score that guides fuzzy aggregation, ensuring anatomically informed and statistically robust decision-making. Experiments were conducted on the SIPaKMeD dataset.

Results

The proposed ensemble achieved strong predictive performance, with accuracy ≈ 0.94, F1-score ≈ 0.94, and area under the curve (AUC) ≈ 0.99. Calibration metrics further confirmed model reliability, with an expected calibration error (ECE) of 0.030, a Brier score of 0.078, and a negative log-likelihood (NLL) of 0.198. The approach consistently outperformed conventional deep learning and fuzzy ensemble baselines.

Conclusions

This study presents an interpretable and reliability-aware fuzzy ensemble framework that advances AI-assisted cervical cancer screening. By integrating explainability alignment and calibrated confidence into a unified reliability measure, the method fosters both diagnostic accuracy and clinical trust, marking a significant step toward safe, transparent medical AI systems. Comparable performance was also observed on an independent external validation dataset, confirming the cross-dataset generalization capability of the proposed framework.