Abstract
Statement of Problem
Cytoplasm and nuclei detection and segmentation of the cervical pap smear images plays essential role in computer-assisted diagnosis of the cervical cancer. Segmentation task of cytology image is obstructed by unclear cytoplasm borders, squamous, overlapping, and folded cells. This project proposes a method to detect nuclei, as well as segment nuclei and cytoplasm of cervical cytology images with the help of U-Net based architecture deep learning models.
Sources of Data
In this study, I employed the Cervix93 and ISBI2014 datasets for training the deep learning models.
Conclusions Reached
During this research, it was noted that both CNN and Vision Transformer models delivered competitive results in the segmentation of cervical cell components, specifically nuclei and cytoplasm. The custom-built Swin U-Net model excels in detecting overlapping segmentations, showcasing exceptional performance. These discoveries indicate the potential application of these models in an image-based computerized analysis system for the early detection of cervical cancer.