Accurate and explainable automated analysis of cervical cytology remains a major challenge for artificial intelligence (AI)-assisted cervical cancer screening, particularly in low-resource settings where access to expert cytotechnologists is limited. Although deep convolutional neural networks (CNNs) have substantially improved classification accuracy, their limited interpretability hinders clinical trust and routine adoption. Conversely, neuro-fuzzy systems provide transparent decision-making but lack the ability to learn complex cytomorphological representations directly from image data. Here, we present PapsAI XNet, a novel explainable hybrid deep learning framework that integrates attention-guided CNN feature extraction with an Adaptive Neuro-Fuzzy Inference System (ANFIS) to simultaneously achieve high diagnostic accuracy and interpretable decision-making. PapsAI XNet incorporates a Channel Attention Module and feature normalization strategy to enhance morphological feature discrimination and stabilize downstream fuzzy inference. A compact ANFIS rule base initialized through data-driven clustering enables transparent reasoning while avoiding the rule explosion commonly associated with high-dimensional neuro-fuzzy systems. The framework was evaluated on the publicly available Herlev dataset comprising seven cervical cytology classes and benchmarked against ResNet-18 and MobileNetV2 using accuracy, sensitivity, specificity, precision, F1-score, receiver operating characteristic (ROC) analysis, computational complexity, and explainability analyses. PapsAI XNet achieved an overall accuracy of 97.8%, sensitivity of 96.4%, specificity of 98.6%, precision of 97.1%, and F1-score of 96.7%, significantly outperforming both benchmark CNN architectures. The greatest improvements were observed for mild and moderate dysplasia, where gradual morphological transitions frequently challenge conventional deep learning models. Class-wise area under the ROC curve exceeded 0.97 for all abnormal cytological categories. Furthermore, learned Gaussian membership functions, fuzzy-rule surfaces, and rule activation patterns provided clinically meaningful explanations linking deep morphological features to diagnostic outcomes while maintaining a lightweight computational footprint suitable for real-time deployment. These findings demonstrate that integrating attention-guided deep feature learning with interpretable neuro-fuzzy reasoning enables accurate, explainable, and computationally efficient cervical cytology classification. PapsAI XNet provides a practical framework for trustworthy AI-assisted cervical cancer screening and represents an important step toward clinically deployable explainable AI for digital cytopathology, particularly in resource-constrained healthcare systems.

Explainable hybrid deep learning for automated cervical cytology classification

期刊论文
出版模式：

开放获取
发表日期：

2026年7月27日

查看全文

Wasswa William,
Andrew Ware

摘要

Accurate and explainable automated analysis of cervical cytology remains a major challenge for artificial intelligence (AI)-assisted cervical cancer screening, particularly in low-resource settings where access to expert cytotechnologists is limited. Although deep convolutional neural networks (CNNs) have substantially improved classification accuracy, their limited interpretability hinders clinical trust and routine adoption. Conversely, neuro-fuzzy systems provide transparent decision-making but lack the ability to learn complex cytomorphological representations directly from image data. Here, we present PapsAI XNet, a novel explainable hybrid deep learning framework that integrates attention-guided CNN feature extraction with an Adaptive Neuro-Fuzzy Inference System (ANFIS) to simultaneously achieve high diagnostic accuracy and interpretable decision-making. PapsAI XNet incorporates a Channel Attention Module and feature normalization strategy to enhance morphological feature discrimination and stabilize downstream fuzzy inference. A compact ANFIS rule base initialized through data-driven clustering enables transparent reasoning while avoiding the rule explosion commonly associated with high-dimensional neuro-fuzzy systems. The framework was evaluated on the publicly available Herlev dataset comprising seven cervical cytology classes and benchmarked against ResNet-18 and MobileNetV2 using accuracy, sensitivity, specificity, precision, F1-score, receiver operating characteristic (ROC) analysis, computational complexity, and explainability analyses. PapsAI XNet achieved an overall accuracy of 97.8%, sensitivity of 96.4%, specificity of 98.6%, precision of 97.1%, and F1-score of 96.7%, significantly outperforming both benchmark CNN architectures. The greatest improvements were observed for mild and moderate dysplasia, where gradual morphological transitions frequently challenge conventional deep learning models. Class-wise area under the ROC curve exceeded 0.97 for all abnormal cytological categories. Furthermore, learned Gaussian membership functions, fuzzy-rule surfaces, and rule activation patterns provided clinically meaningful explanations linking deep morphological features to diagnostic outcomes while maintaining a lightweight computational footprint suitable for real-time deployment. These findings demonstrate that integrating attention-guided deep feature learning with interpretable neuro-fuzzy reasoning enables accurate, explainable, and computationally efficient cervical cytology classification. PapsAI XNet provides a practical framework for trustworthy AI-assisted cervical cancer screening and represents an important step toward clinically deployable explainable AI for digital cytopathology, particularly in resource-constrained healthcare systems.