CerviGAN: Cervical cancer pap smear image classification with GAN-based data augmentation and external attention transformer

Cervical cancer remains a significant global health concern among women. Timely detection and diagnosis are critical to reducing mortality and morbidity rates. Although the Papanicolaou (Pap) smear test is widely employed for screening, it is time-consuming, labor-intensive, and prone to human errors due to its subjective nature. Artificial intelligence (AI) techniques offer promising alternatives; however, challenges such as limited datasets and sparse annotated images hinder the development of robust automated diagnostic models. In this study, we propose CerviGAN, a DCGAN-EAT methodology that integrates a deep convolutional generative adversarial network (DCGAN) with an external attention transformer (EAT). DCGAN is utilized to address class imbalance by generating high-quality synthetic images of the minority class, while the EAT model is employed for automated classification by capturing crucial global features. Our approach effectively mitigates dataset imbalance and enhances feature representation through external attention mechanisms. Extensive ablation studies further demonstrate that the proposed model outperforms a standard Vision Transformer without external attention, validating the architectural improvements. The external attention block enables better modeling of interactions across diverse samples, strengthening the model’s ability to capture global dependencies. Experimental evaluations reveal that our model achieves a notable 3-class classification accuracy of 98.53% on the publicly available Sipakmed dataset, surpassing or matching state-of-the-art methods. These results underscore the effectiveness of CerviGAN in classifying cervical Pap smear images and its potential applicability for improving clinical diagnosis workflows.