Robust Segmentation of Overlapping Cells in Cervical Cytology Using Light Convolution Neural Network

Abstract

Automated segmentation of cells in cervical cytology images poses a great challenge due to the presence of fuzzy and overlapping cells, noisy background, and poor cytoplasmic contrast. We present an improved method for segmenting nuclei and cytoplasm from a cluster of cervical cells using convolutional network and fast multi-cell labeling. A light convolutional neural network (CNN) model is employed to generate nuclei candidates, which can serve as accurate initializations for the subsequent level set segmentation and provide a priori knowledge for the cytoplasm segmentation. A fast multi-cell labeling method based on the superpixel map is devised to roughly segment clumped and inhomogeneous cytoplasm before applying a cell boundary refinement approach. A shape constraint in conjunction with boundary and region information drive a level set formulation to perform a robust cell segmentation. The qualitative and quantitative evaluations demonstrated that the presented cellular segmentation method is effective and efficient.