Research On Automatic Segmentation Of Nuclei In Pathological Images

Pathological diagnosis is the final diagnosis of the disease.Pathology is the only way to diagnose the disease when other medical images cannot be confirmed and the accurate segmentation of the nucleus is the basis of pathological diagnosis.Artificial nucleus segmentation and disease assessment are extremely time-consuming and the results of different expert assessments can also produce subjective differences.Therefore,automatic nucleus segmentation has gradually become the focus of research in recent years.The task of automatic segmentation of nuclei is facing great challenges due to the interference of different tissues,different solubility of dyes,different dyeing time and overlap of nuclei in pathological images.The color of pathological image is normalized to make the color distribution of all images the same while maintaining the image structure unchanged in order to reduce the difficulty of segmentation caused by color difference caused by staining.Two methods are proposed to segment the foreground.The first one is based on the geodesic active contour model.The rough segmentation of nucleus is obtained by adaptive threshold method and refined by level set evolution.Another method is based on the deep learning.The nuclei is accurate segmented without reducing image resolution which using residual structure and dilated convolution to extract image feature information.Pathology is based on the morphology of a single nucleus for diagnostic analysis,but there will be overlap of nuclei in pathological images.The overlapping nuclei need to be splited after obtaining the foreground of nuclei.The overlapping regions are separated by convexity analysis according to the hypothesis that the single nucleus has convex structure and the convexity of overlapping nuclei is destroyed.Then the concave points are detected to obtain the intersection points of different nuclei,so as to split the overlapping nuclei.The method is tested on three different data sets and the qualitative and quantitative analysis of the proposed method is carried out.The Dice scores of the two methods are 0.81,0.801 and 0.845,0.821 respectively on the open data sets.The experimental results show that the proposed method can accurately segment the nucleus foreground and effectively solve the problem of overlapping nuclei.