A Omni-Semantic Aware Transformer for Cervical Cytopathology Screening

Area of Science:

• Pathology
• Computer Vision
• Medical Imaging

Background:

• Cervical cell image classification methods often overlook fine morphological details.
• This underutilization limits the accuracy of pathological cell image screening.

Purpose of the Study:

• To propose an omni-semantic aware transformer (OSAT) for enhanced cervical cell image classification.
• To improve the underutilization of fine morphological details in pathological cell image screening.

Main Methods:

• OSAT integrates image patch embedding with a multi-head convolutional block for multi-scale semantic extraction.
• A novel omni-semantic feature learning module enhances interactions between shallow and deep features.
• The model was evaluated on the SIPaKMeD dataset and a new 15-category dataset.

Main Results:

• OSAT demonstrated superior accuracy and robustness in cervical cell image classification.
• The proposed method effectively captures complementary structural and pathological semantics.
• Enhanced utilization of cell boundaries and textural patterns was observed.

Conclusions:

• OSAT offers a significant advancement in pathological cell image screening.
• The model's ability to integrate multi-level semantic features improves diagnostic accuracy.
• OSAT provides a more robust approach compared to existing state-of-the-art models.