Selective Detection and Segmentation of Cervical Cells

Abstract

Accurate detection and segmentation of cervical cells is often considered as a critical prerequisite of the prediction of dysplasia or cancer either by a pap smear or the lately developed liquid-based cytology (LBC). The computer-aided detection in microscope images can relieve the pathologists from strenuous manual labors with higher accuracy and efficiency. In the segmentation tasks of real-life clinical data, one challenging issue is the mis-identification of other cells, such as inflammatory cells, with similar appearance of nuclei in shape, size and texture. With a large distribution in the whole slide, even overlap up to 50% to 75% percentage of normal or abnormal cells, these cells are usually detected and segmented as nuclei. In this paper, compared with the typical three-catalogue segmentation methods of nuclei, cytoplasm and background proposed in the literature, we provide a discrimination between inflammatory cells and nuclei by adding a new catalogue. We present two novel convolutional neural networks (CNN), a deeply fine-tuned model and a trained from scratch model. The models enable us to sensitively detect and remove background noises such as mucus or red blood cells. We also profile a detailed performance comparison between these two methods, with the advantages of either network presented. The experiments are based on the sufficient clinical dataset we collected, and the results show the effectiveness of proposed approaches in selective cell detection and segmentation.