Adversarial Mask Generation for Robust Cervical Cell Segmentation

Accurate segmentation of cervical cells is critical for the early diagnosis and treatment of cervical cancer. However, overlapping and touching cells, as well as variations in color and texture caused by different collection techniques, present significant challenges for automated analysis systems. This study proposes a robust and adaptable cytoplasm segmentation method based on a generative adversarial network to address these challenges. The proposed approach effectively detects individual cytoplasm regions in cervical cell images with varying characteristics and is evaluated using three publicly available datasets: Herlev, ISBI 2014, and CISD. Two experimental scenarios are conducted to assess the model's generalization and adaptability. In the first scenario, the model is trained on a combined dataset comprising samples from all three datasets and tested on each dataset individually. In the second scenario, the model is trained on a single dataset and tested on previously unseen datasets to evaluate its generalization ability. Qualitative and quantitative results from both scenarios demonstrate that the proposed method achieves high segmentation accuracy and successfully adapts to diverse image types. These findings suggest that the method is a promising candidate for integration into automated cervical cancer screening systems.