Research On Abnormal Cell Segmentation Method Based On Lung Pathological Image

Compared with medical images such as CT and ultrasound,pathological images can provide more effective diagnostic information,therefore they are widely used in clinical diagnosis and medical research.In the traditional diagnosis of lung cytopathology,the shortage of pathologist resources,the complexity of abnormal cell morphology and the arrangement of different lung diseases lead to the increasing rate of misdiagnosis at diagnosis of lung diseases.Therefore,the image-assisted diagnostic system of lung cytopathology,which helps to reduce the misdiagnosis rate,has very important research significance.However,the accurate segmentation of lung abnormal cells,as the most critical step in this system,is not well realized in the existing researches.Therefore,this thesis focuses on the pathological images of lung cells to deeply study the segmentation of abnormal cells.The main research work includes the following aspects.In the image preprocessing stage,firstly,aiming at the problem of gray impregnated plaque noise in the image,the combination of image opening operation and image restoration algorithm is used to denoise the image;secondly,focusing on the issue of unbalanced brightness in the image,the image brightness enhancement algorithm based on low illumination image enhancement is utilized to increase the contrast between the target area and the background area in the lung pathological image;finally,for the problem of the small number of lung cell images,the method of image geometric transformation is addressed to complete the data enhancement processing.In the segmentation stage of abnormal lung cells,we use three network models with codingdecoding structure to implement the segmentation task of abnormal lung cells,and then two evaluation indexes of IOU and DSC are used for the performance evaluation of each model.The experimental results show that the IOU and the DSC of the U-Net model are 0.6519 and 0.7743 respectively.And both evaluation indexes of U-Net are higher than the other two segmentation models,indicating that the segmentation performance of the U-Net model is relatively better.In order to further improve the segmentation accuracy of the model,we propose a segmentation model of abnormal lung cells based on the improved U-Net.Firstly,aiming at the difficulty of feature extraction in some small target regions of abnormal cells,a dense connection mechanism is added between convolutional layers of the U-Net encoder to improve the propagation ability between features and extract more feature information of abnormal cells;secondly,addressing to the problems of uneven brightness from abnormal lung cell image and difficult accurate segmentation of abnormal cell contour,an attention mechanism is added to the bottleneck layer of the U-Net to increase the weight of abnormal cell area and avoid the interference of image brightness imbalance to the model;finally,due to the initial segmentation result of the model is not refined enough,the fully connected conditional random field algorithm is utilized to refine the segmentation result to improve the accuracy of the segmentation result.The experimental result shows that the IOU and DSC of the improved U-Net are respectively0.7149 and 0.8211.Compared with the U-Net model,the IOU and DSC of the proposed model are increased by 9.66% and 6.04% respectively,indicating that the model is not only more suitable for the segmentation task of abnormal lung cells,but also better distinguish low-contrast areas,small target areas,and abnormal cells with diverse shapes.