An Explainable Multi-Level Framework for Cervical Cancer Detection Using Traditional Computer Vision and Deep Learning

Cervical cancer remains one of the leading causes of cancer-related mortality among women worldwide, highlighting the need for accurate and early detection methods. In this study, we propose an explainable multi-level framework for cervical cancer detection by combining traditional computer vision techniques with advanced deep learning models. The framework leverages the SIPAKMED Pap smear dataset, which provides high-quality annotated images for multi-class classification. Three novel architectures are developed and evaluated: PSMF‑Net (Pap‑Swin Morpho‑Focus Net), designed to capture multi-scale morphological features using EfficientNetV2B3 enhanced with Squeeze-and-Excitation blocks; CAD‑Fusion (Cervi‑Attention Dual‑Path Fusion), which employs a dual-path attention mechanism to integrate coarse lesion context and fine cellular details without computational overhead; and KCPD‑Net (Kernel‑Constrained Perinuclear Dispatcher Network), which focuses on perinuclear regions with custom kernel-constrained convolutions to extract fine-grained features. The framework incorporates pre-processing, multi-level feature extraction, and fusion, followed by classification into standard Pap smear categories, including. Explainable AI techniques such as Grad-CAM and SHAP are applied to visualize the regions contributing to predictions, enhancing interpretability. Experimental results demonstrate the effectiveness of the proposed models in capturing relevant morphological patterns and improving classification accuracy, providing a robust and interpretable solution for automated cervical cancer screening.

Keywords: Cervical Cancer Detection, Pap Smear, PSMF‑Net, CAD‑Fusion, KCPD‑Net, Explainable AI, Multi-Level Feature Extraction, Deep Learning, Computer Vision.