Exploring Data Imbalance Challenges in Cervical Cancer Detection Using Advanced Deep Learning Models

Cervical cancer presents a significant global health challenge, ranking as the fourth most widespread cancer world-wide. It is originated in the cervical epitheliumis and is primarily caused by persistent infection with high-risk strains of the human papillomavirus (HPV). Traditional screening modalities such as Pap smears and visual inspections have long been employed for detection of cervical cancer; however, recent advancements in computational techniques and image-based classification methodologies, integrating machine learning (ML) and deep learning (DI), have introduced a new era of precision diagnostics. A common concern for researchers is the issue of balanced data and its impact on the accurate classification of cervical cells. This study emphasizes this concern by implementing a set of DL techniques on two distributions of the SIPaKMeD Cervical Cancer Dataset: one is a balanced distribution with five classes, and the other is unfavorably uneven binary classification. The DL strategies encompass five ConvNets including VGG-16, DenseNet-121, ResNet-50 V2, EfficientNetV2 BO, and ConvNeXtBase, as well as Vision Transformer (ViT-16) and a multilayer CNN trained from scratch, utilizing transfer learning (TL) and fine-tuning (FT) to investigate the models' behavior on each distribution. For the five-class distribution, a testing accuracy of 100% was accomplished by fine-tuning DenseNet-121 and EfficientNetv2 B0, while the highest accuracy reached in the two-class approach was a 98.09% that was also achieved by the fine-tuning of DenseNet-121.