Light-XAI: a CADx for explainable cervical cancer detection via attention-based lightweight convolutional neural networks and layer-wise feature fusion.

Cervical cancer, the fourth most common cancer, requires prompt identification for treatment. The Pap smear test is the standard for cervical cancer detection and diagnosis, but it is complicated and laborious, and its efficacy depends on medical professionals. Computer-aided detection and diagnosis (CADx) systems employ deep learning, especially convolutional neural networks (CNN), to improve the precision and efficiency of cervical cancer diagnosis. Nevertheless, conventional CNNs do not inherently utilise instinctive methods of evaluating the relevance of features, as a qualified medical professional would, and instead assess features more broadly. In addition, CNNs are extremely efficient at extracting spatially invariant local features. Self-attention mechanisms allow CNNs to prioritise semantically significant clinical feature regions, gather relevant information with long-distance connections, and could efficiently model global-level features. To this end, this study integrates the self-attention mechanism into a CNN architecture and presents “Light-XAI”, a comprehensive lightweight explainable AI (XAI) CADx system for cervical cancer classification from Pap smear images. The system employs combinations of lightweight CNNs and applies the local gradient-weighted class activation mapping (Grad-CAM) technique for the explainability of decisions made. Furthermore, deep features from each CNN trained with Pap smear images are derived from two deep layers opposing current CADxs which either obtain features from a single layer or perform end-to-end CNN classification. In addition, features extracted from both deep layers of self-attention CNN are combined and decreased using a feature reduction method. Furthermore, the attributes of the three self-attention CNN models are fused, and subsequently a feature selection method is employed to choose the most impactful features. The results demonstrate that Light-XAI has achieved an accuracy of 97.27% and 99.82% for the SIPaKMeD and Mendeley liquid-based cytology (LBC) datasets, respectively, suppressing most of the present CADx. This superior performance allows Light-XAI to serve as a powerful tool to categorise cervical cancer. In addition, it shows the potential for improving diagnostic precision, allowing for the adoption of advanced methods in the field of medical image analysis.