Hybrid Deep Learning Framework for Cervical Cancer Cell Classification Using Discrete Wavelet Transform and EfficientNetV2-B3

Abstract

Cervical cancer is among the most common and fatal cancers that affect women worldwide, timely and accurate diagnosis using cytopathology is essential in determining treatment options. Traditional cytological tests, including the Pap smear, involve manual processes that are laborious, time-consuming, and highly variable between different observers. In this paper, a novel dual-path hybrid deep learning architecture is proposed for automatic multi-class classification of cervical cancer cells based on the integration of Discrete Wavelet Transform (DWT) as frequency-domain features with deep spatial features extracted using the EfficientNetV2-B3 architecture. The proposed architecture extracts multi-scale wavelet features of dimension 256 in parallel with deep semantic features of dimension 1280. Feature vectors derived from both pathways are concatenated to create an integrated 1536-dimensional feature representation for multi-class classification of cervical cancer cells into five categories. The result is compare two experimental designs in the study: (1) EfficientNetV2-B3 alone with 97.04% accuracy and 0.9908 AUC score, and (2) hybrid Deep Wavelet + EfficientNetV2-B3 model with higher performance results (accuracy: 99.26%, precision: 99.28%, recall: 99.26%, F1-score: 99.27%, and AUC: 99.38%).