A Multi-Representation Hybrid CNN Feature Extraction Framework for Cervical Pre-Cancer Image Classification

Abstract

Cervical cancer remains a major cause of mortality among women, particularly in low-resource settings, highlighting the need for reliable image-based screening methods. Visual Inspection with Acetic Acid (VIA) is widely used in clinical practice; however, its diagnostic accuracy strongly depends on operator experience and subjective interpretation. This study systematically evaluates a multi-representation feature extraction framework for cervical pre-cancer image classification by comparing three paradigms: conventional handcrafted features, convolutional neural network (CNN)-based deep features from pre-trained networks, and hybrid features combining both representations. Experiments are conducted on three image representations, namely original VIA images, Frangi-filtered images, and morphology-based vessel images, using AlexNet, ResNet-50, and EfficientNet as deep feature extractors. All feature representations are classified using KNN and SVM under a strict evaluation protocol. The results show that hybrid feature extraction performs best on original images, with HybridAlexNet achieving an accuracy, sensitivity, and specificity of 0.84. The findings indicate that representation-aware hybrid feature design provides a robust and interpretable solution for screening-oriented cervical pre-cancer detection.