Automated Segmentation of Overlapping Cells in Cervical Cytology Images Using Deep Learning

In this paper, we deal with segmentation of overlapping cells in cervix cytological images. One of the major challenges in analysing cervical cytology images is the segmentation of overlapping cells. The complexity is mainly due to low contrast, wide range of overlapping ratios and other debris. We propose to address these issues using deep learning. Our method is a three-stage process and uses the convolutional network U-Net in two of these stages. We showcase the process with the Overlapping Cervical Cytology Image Segmentation Challenge dataset. Innovations by using available information provide results that are more promising than existing methods. These include the use of overlap percentage, and ensuring that the inputs to the second U-Net are the original image, the cell mass, and the masks of the individual nucleus.