Temporal-Aware Multi-Image Vision Transformer for Cervical Cancer Screening via Cervicography

Cervical cancer remains a critical public health challenge in low- and middle-income countries (LMICs), where AI-assisted cervicography offers promising potential for scalable screening solutions. However, current AI-based single-image classification models face a fundamental limitation: when applied to multiple tele-cervicography images from the same patient, they produce inconsistent predictions that create diagnostic uncertainty and undermine clinical confidence. To address this challenge, we present a temporal-aware, multi-image Vision Transformer (ViT) architecture that processes multiple cervical images collectively, generating unified patient-level predictions using a five-class framework (Negative, Atypical, Positive 1, Positive 2, Positive 3). Our approach incorporates temporal embeddings to capture sequential relationships between images acquired during visual inspection with acetic acid (VIA) examinations, enabling the model to leverage complementary information across the image sequence. Experimental results demonstrate that our multi-image ViT-Huge variant achieves 70.4% accuracy when processing all patient images, outperforming single-image baselines and establishing temporal-enhanced Vision Transformer as an effective solution for reliable AI-assisted cervical cancer screening in resource-limited settings.