A study on semi-supervised solutions for overlapped cell classification

Abstract

This paper proposes a semi-supervised learning-based approach to segment and classify overlapping cervical cells. The segmentation model is based on DoNet, which effectively separates intersecting cell regions using a decompose-and-recombine strategy. For classification, we apply a Semi-FixMatch model that leverages unlabeled data with pseudo-labeling and consistency regularization. Experiments on the ISBI2014 dataset show that our method achieves competitive performance even with limited labeled data, accurately identifying abnormal cells based on the nucleus-to-cytoplasm (N/C) ratio. The proposed solution enhances the reliability of automated cytology analysis and reduces the burden of manual labeling.