Research And Application On Cell Segmentation Based On Deep Learning

Cellular immunohistochemistry is a commonly used method for cancer screening,which uses antibody-conjugated nano-fluorescent probes to label antigen expression in cancer cells and obtain pathological images through electron microscopy.Pathologists can preliminarily screen cancerous cells by analyzing the amount of fluorescence probes entering the cells in pathological images.However,it is difficult to visually judge the amount of fluorescence entering of the cells because of the dense distribution,diverse shapes,adhesion or overlap of cells,and the blurry boundary with the background,it is also challenging to quantify,which requires a significant amount of time and effort from doctors.Therefore,this study focuses on the segmentation and recognition of cell pathology images,and designs a lung cancer auxiliary diagnosis system to statistically analyze the average fluorescence intensity of cells in pathological images and assist doctors in selecting suspicious cells through delivery methods to assist doctors in diagnosing lung cancer.The research work mainly includes the following three aspects:(1)This study introduces Elliptical Fourier descriptors to describe the contours of cells.Elliptical Fourier descriptors approximate the boundary curve of the target object using frequency wave superposition.It represents the two-dimensional coordinates of contour points with Elliptical Fourier descriptor parameters and has good performance in describing the contour of elliptical regions.This study converts the boundary pixel labels of cells into contour coordinates,and then transforms them into Elliptical Fourier descriptor parameters in the frequency domain to represent cell contours.The experiment shows that elliptical Fourier descriptors have the minimum representation error and good linear space continuity in representing cell contours,which is suitable for describing cell contours.(2)A cell contour recognition network based on elliptical Fourier descriptors and feature pyramid structure is proposed.The network uses elliptical Fourier descriptors to represent cell contours and predicts a set of elliptical Fourier parameters to recognize the position and shape of cells.The network uses a feature pyramid structure in the contour recognition module to recognize contours at multiple scales of large,medium,and small sizes,solving the problem of difficult training and inaccurate prediction due to uneven cell contour sizes.In comparative experiments on multiple cell datasets,the proposed method in this study has an average F1 score of 0.658 on the DSB2018 dataset and an average F1 score of 0.480 on the MoNuseg2018 dataset,both of which are better than other methods,demonstrating the effectiveness of the proposed model.(3)A lung cancer auxiliary diagnosis system is designed for the screening of cellular immunohistochemistry pathological images of lung cancer.Combining a novel fluorescent probe imaging method to obtain pathological images of lung epithelial cells,computer technology and the cell instance segmentation algorithm proposed in this study are used to segment and detect cell images.The system realizes the automatic segmentation of cell images and the identification of suspicious cells to assist expert doctors in clinical diagnosis.The effectiveness of the system in identifying suspicious cells is verified through experimental validation using a dataset of sample images.