XQE-Net: A deep learning framework with integrated feature fusion and optimization for cervical cancer screening

Abstract

Cervical cancer remains a critical global health concern requiring accurate and efficient automated screening solutions. This paper presents XQE-Net, a Computer-Aided Diagnosis framework built upon a heterogeneous ensemble of three CNN architectures — Xception, SqueezeNet, and EfficientNet B7 — each selected for its distinct architectural design. XQE-Net extracts hierarchical deep features from three layers of each CNN — fully connected, pooling, and convolutional — yielding nine independent feature sets that collectively capture a comprehensive diagnostic representation of cervical cell morphology. A Discrete Wavelet Transform using a Haar basis function at six decomposition levels is applied exclusively to convolutional layer features, compressing raw dimensionality by up to 98.9% while preserving discriminative spectral characteristics prior to fusion. Subsequently, mRMR-based feature selection reduces the fused feature space by up to 91.6%, yielding a deployment-ready classifier requiring as little as 1.56 KB memory per sample. XQE-Net process Pap smear images without requiring segmentation or image enhancement, simplifying the diagnostic pipeline for clinical deployment. Validated through five-fold cross-validation on two benchmark datasets, XQE-Net achieves 98.5% accuracy on SIPaKMeD and 100% on Mendeley LBC, outperforming existing methods by 3.77% and 1.22% respectively, confirming its diagnostic reliability and cross-dataset generalizability.