A Omni-Semantic Aware Transformer for Cervical Cytopathology Screening.

Abstract

Traditional cervical cell image classification methods primarily transform shallow semantic cues into deep semantic representations, often leading to the underutilisation of fine morphological details, such as cell boundaries and textural patterns. To address this limitation, we analysed the correlations among multi-level semantic features in cervical cell images and proposed an omni-semantic aware transformer (OSAT) for pathological cell image screening. Specifically, OSAT employs a multi-scale semantic extraction architecture that integrates image patch embedding with a multi-head convolutional block, thereby capturing complementary structural and pathological semantics across different receptive fields. Furthermore, we designed a omni-semantic feature learning module to explicitly model and enhance the interactions between shallow morphological features and deep pathological representations. The proposed method was evaluated on the publicly available SIPaKMeD dataset and a newly constructed dataset containing 15 categories of cervical cytopathology images. Extensive experimental results demonstrate that OSAT achieves superior accuracy and robustness compared with current state-of-the-art classification models.