Benchmarking Image Models Including CNNs, Transformers, and Hybrid Architectures for Cervical Cell Classification

Cervical cancer screening is a critical component of women's health, requiring accurate and efficient analysis of Pap smear cell images. In this study, conducted as part of the Pap Smear Cell Classification Challenge (PS3C) at IEEE ISBI 2025, we benchmark a variety of modern image architectures, including Convolutional Neural Networks (CNNs), Vision Transformers, and hybrid models, to classify cervical cell images into Healthy, Unhealthy, and Rubbish categories. Our systematic evaluation of seven models reveals that hybrid architectures, particularly MaxViT (F1=0.86633), demonstrate superior performance, though the margin across different architectures remains relatively small. We achieved 2nd place (F1=0.86797) in the development phase using ensemble strategies and 5th place (F1=0.79351) in the final phase with a single MaxViT model among 41 and 16 teams respectively. By maintaining comparable parameter counts (20M-30M) across all models, our study provides a fair comparison of architectural choices while considering practical deployment constraints in clinical settings.