Research On Segmentation Method Of Cervical Cell Image

Cervical cancer is a highly prevalent cancer affecting women’s health,and the World Health Organization has proposed a global action plan to eradicate cervical cancer.Therefore,it is imperative to improve the efficiency of cervical cancer screening and diagnosis with the help of computer technology.Segmentation of cervical cell images is one of the key tasks of the cervical cancer-assisted diagnosis system,which can facilitate pathological screening and clinical diagnosis,so segmentation of cervical cell images is of great significance in medical and scientific fields.Although a lot of work has been done to improve the accuracy of cervical cell image segmentation,its segmentation accuracy is still limited by some common problems,and this thesis focuses on the following problems:(1)image impurities and noise interference,color inconsistency,and blurred details;(2)irregular cell cluster boundaries and complex spatial information at overlapping concave points;(3)non-epithelial cell components interfere with cell nuclei segmentation;(4)multiple cell overlap.To address the above problems,this thesis investigates the cervical cell image segmentation method using deep learning methods.For cervical cell image preprocessing,non-local mean filtering is used to remove fine impurities and noise interference in the image;for the problems of uneven staining and poor contrast in the production,a multi-scale Retinex algorithm with color recovery is used for color correction;for the problem of blurred medical image edges,multi-scale bilateral Gaussian filtering is used for edge enhancement.In cervical cell cluster segmentation,a cell cluster segmentation method based on multiscale feature fusion with an attention mechanism is proposed.Based on the Deep Labv3+network,a feature pyramid structure is used to fuse the mid-level features of Res Net50,and the hybrid attention idea is introduced.According to assigning different weights to the highlevel semantic features and low-level detail features,the model is enhanced to learn local detail information and avoid the interference of neutrophils attached to the edges of cell clusters to complete the segmentation of cell clusters in cervical cell images.For cervical cell nuclei segmentation,a cell nuclei segmentation method based on twoway features is proposed.By fusing the structures of Res Net50 and Transformer in parallel as the backbone network,and using the two-way feature fusion module to fuse the two branches of features.The cervical cell nuclei segmentation method based on Transformer’s self-attentive mechanism and fused two-way features not only facilitates to emphasis of contextual information with global dependence but also improves the recognition of cell nuclei by the model.For the problem of multi-cell overlap in cervical cell clusters,a post-processing method based on the active contour model is proposed.The initial contours of cells are delineated by the segmentation results of cell clusters and nuclei,and the curvature information of the contours is introduced to construct the active contour model to realize the multi-contour evolution and obtain the refined overlapping cervical cell boundaries.The experimental results of the proposed cervical cell image segmentation method illustrate that it successfully achieves highly detailed segmentation of both cervical cell masses and nuclei while minimizing the impact of interference information on the segmentation results.The MIo U scores for cell mass and cell nucleus segmentation are respectively 0.9178 and 0.8329,which are 2.14% and 2.33% higher than the baseline model.Additionally,the proposed method also accurately segments overlapping cells and obtains the MIo U score of0.9083 for overlapping cell contours.