Reliable Cervical Cancer Screening Using Cytology Data Through An Histopathology Foundation Model

Cervical cancer remains a leading cause of mortality among women, underscoring the need for accurate and reliable screening tools. While deep learning (DL) mod-els have shown promise in cervical cancer classification, most approaches over-look the crucial aspect of uncertainty quantification, which is essential for trust-worthy predictions in clinical settings. In this study, we explore the application of the UNI histopathology foundation model to cytology data, comparing its per-formance with two states of the art convolutional neural networks (CNNs): Res-Net50 and DenseNet121. We assess each model on classification performance, calibration and their uncertainty as well as their ability to represent class separa-bility using t-SNE visualizations. Our results demonstrate that the UNI model significantly outperforms ResNet50 and DenseNet121, achieving near-perfect ac-curacy and calibration. t-SNE plots further reveal that the UNI model produces well-defined, distinct clusters, indicating superior feature representation. Fur-thermore, the analysis of predictive entropy shows that the UNI model exhibits lower uncertainty in well-classified samples and effectively distinguishes between correct and incorrect predictions, reinforcing its reliability. These findings suggest that the UNI model is a highly effective and reliable tool for cervical cancer screening, paving the way for future advancements in uncertainty-aware diagnos-tic applications and the adoption of Histopathology model in Cytopathology.