Research On Cell Image Segmentation Based On Deep Learning

Image segmentation is particularly important in cervical cancer automatic screening technology,which determines the accuracy of doctors’ final discrimination.So far,many scholars have proposed a lot of methods to solve the problem of cervical cell image segmentation.However,cell segmentation has always been a difficult problem due to many complex interference situations in cell images,especially for overlapping cell segmentation,it is difficult to take into account a variety of complex situations using traditional methods.To solve these problems,this paper proposes a new method for cell segmentation based on deep learning.Firstly,the nucleus is defined as the localization factor to assist in the segmentation of overlapping cells.The CE-Net network is used to extract the localization factors accurately,so as to eliminate the influence of the shape of the nucleus on the cytoplasmic segmentation.Combined with the migration learning and Res Net algorithm,the localization factors are classified into naked cells and normal cells.Then,a new cell segmentation model SEG-GAN is improved based on the generative confrontation network to segment cells in Generative way;at the same time,the model loss is defined from two aspects of generative task and segmentation task to ensure the segmentation quality of SEG-GAN model.Finally,a contour fitting algorithm based on level set is proposed to extract the cell contour after SEG-GAN model segmentation as the initial contour by minimizing the corresponding energy functional of each cell,the optimized cell contour is obtained and the final cell segmentation is completed.Through the comparative experiment of location factor extraction on self-built data sets,CE-Net network shows the most superior segmentation performance compared with other methods,which can achieve the accurate extraction of location factor.Compared with other excellent traditional methods,SEG-GAN model shows good segmentation ability for overlapping cells,which has certain research value and significance.