Vision Transformer-based classification of cervical cancer images

Cervical cancer is a significant health concern as it is one of the most fatal causes of cancer deaths in women. Detection at an early stage is vital, but conventional diagnostic procedures via Pap smears relies heavily on manual cytopathology inspection, a process prone to human fatigue, diagnostic subjectivity, and severe practitioner shortages in low-resource medical environments. To address these challenges, this study presents an automated cervical cancer image classification framework utilizing advanced Vision Transformer (ViT) architectures. Three ViT variants: Data-efficient Image Transformer (DeiT), Swin Transformer, and Cross-Attention Multi-Scale Vision Transformer (CrossViT) were fine-tuned and evaluated on the benchmark SIPaKMeD dataset comprising five distinct cervical cell classes: superficial-intermediate, parabasal, koilocytotic, dyskeratotic, and metaplastic. Data balance was maintained using a dynamic WeightedRandomSampler combined with image augmentations. Experimental results demonstrate that DeiT achieved the highest classification performance with an accuracy of 95.22% and a weighted average F1-score of 0.9522, outperforming CrossViT (94.82%) and Swin Transformer (93.83%). DeiT was the most accurate with 95.22%, and Swin Transformer and CrossViT clocked 93.83% and 94.82%, respectively. The optimal model was integrated into an interactive web-based interface built with Streamlit, demonstrating single-image inference latency of ~38 ms on GPU and ~142 ms on CPU. These findings confirm that Vision Transformers provide a highly accurate and scalable decision-support mechanism for automated cervical cancer screening. The work in the future will include expanding the pool of images in the dataset with a variety of cervical images, resolving class imbalance by employing methods such as GAN-based augmentation, and using hybrid architectures to improve results. Additional confirmation on other sample databases will also be required to ascertain the robustness of the system in real world solutions.