Segmenting overlapping cervical cell in Pap smear images

Accurate segmentation of cervical cells in Pap smear images is an important task for automatic identification of pre-cancerous changes in the uterine cervix. One of the major segmentation challenges is the overlapping of cytoplasm, which was less addressed by previous studies. In this paper, we propose a learning-based method to tackle the overlapping issue with robust shape priors by segmenting individual cell in Pap smear images. Specifically, we first define the problem as a discrete labeling task for multiple cells with a suitable cost function. We then use the coarse labeling result to initialize our dynamic multiple-template deformation model for further boundary refinement on each cell. Multiple-scale deep convolutional networks are adopted to learn the diverse cell appearance features. Also, we incorporate high level shape information to guide segmentation where the cells boundary is noisy or lost due to touching and overlapping cells. We evaluate the proposed algorithm on two different datasets, and our comparative experiments demonstrate the promising performance of the proposed method in terms of segmentation accuracy.