Optimizing Residual Networks for Cervical Cancer Screening: A Study on the Pap Smear Cell Classification Challenge

This study presents a comparative analysis of deep learning architectures, with a particular emphasis on Residual Network (ResNet) variants, for automated cervical cancer cell detection using the APACC dataset. Addressing the challenge of class imbalance, we explore two distinct strategies: (i) reclassifying images containing both healthy and unhealthy cells as unhealthy and (ii) excluding these ambiguous samples from the training set. Our experimental findings indicate that the latter approach yields superior classification performance. We conduct an extensive evaluation of ResNet architectural variants, including SE-ResNet, ResNeXt, SE-ResNeXt and Wide ResNet. Among standalone models, Wide ResNet-50 demonstrates the highest performance, achieving an F14-score of 0.865 on the test set. Furthermore, we implement an ensemble of the top five performing models, which achieves the best overall performance. These findings underscore the efficacy of advanced deep learning architectures and targeted class balancing strategies in enhancing automated cervical cancer screening systems.