Research On Cell Recognition In Pathological Image Based On Context

With the rapid development of computer technology and the influence of domestic policy inclination,the use of artificial intelligence and image processing technology to assist medical diagnosis research has received more and more attention.In medical clinical examination,the task of cell recognition in pathological images,namely cell location and classification,has important value in the diagnosis of many diseases.Artificial recognition of cells is slow and inefficient,and the histological and cytological information contained in the pathological image itself is very complicated.Therefore,the automatic identification method of cells in pathological images has high practical value.There are many cell identification methods,but these methods do not fully consider or use contextual information,that is,the characteristic statistical information of the pathological image itself and the spatial relationship information between cells.This information will greatly affect the accuracy of cell recognition.Based on this,this paper designs and implements two pathological image cell recognition algorithms from the perspective of context: one is to promote the improvement of cell recognition accuracy by introducing multi-scale image information,and to input pathological images of different scales in pairs to make the model match The spatial topology information in the pathological image and the cell morphology and texture information are weighed and the category judgment is made;the other method is to introduce the cell recognition method of regional information to help the model learn the distribution relationship between the cells of the same category to promote the cell recognition.The main research work of the article is as follows:1.Aiming at the problems of too small cell morphology,dense cell distribution,and uneven cell number in pathological images,an end-to-end network MACE-CNN(Multiple Aggregation and Context Encoding-Convolutional Neural Network)based on multi-scale image input and cell number statistics is proposed:(1)In order to effectively combine multi-scale image information,MACE-CNN has designed a dual-input feature aggregation module,which takes two pathological images with different magnifications in the same region as input,and extracts multi-scale feature representations from them.The superposition of the product operation performs feature aggregation in real time,and merges the spatial distribution information of the cells in the image with the cytological information such as the texture and morphology of the cells.(2)Aiming at the problem of the uneven number of cells in pathological images,MACE-CNN introduces an attention mechanism module to selectively highlight specific types of cells in the image to reduce the recognition interference caused by cell types that do not exist in some images.(3)Aiming at the problem of dense cell distribution and small morphology,improved based on the mean square loss error,separated the background and the cell area,and gave the cell area a higher loss weight to enhance the accuracy of the model for cell positioning.Compared with previous methods,MACE-CNN is more inclined to complement global and local information in the process of cell recognition.By introducing multi-scale pathological images and cell number information in the recognition process,it can produce better Recognition effect.2.Aiming at the problem of fuzzy distinguishing characteristics of cells between different categories in pathological images and large differences in cell morphology between the same categories,using the characteristics of the clustering of cells of the same category,a pathological image cell recognition method based on local area constraints,Tri-NET,is designed.Through the coordinated task of the cell position and the same cell ROI(Region of Interest,region of interest),the cell classification accuracy is improved.(1)Aiming at the problem of excessive differences between cells in the same category,the classification of cells in the same area is restricted by learning the ROI of the same cells,so that the cells in the same area tend to be classified into the same category.(2)In order to prevent the background area from interfering with cell recognition,the cell location branch is introduced to reduce the wrong location of cells in the background area.(3)Aiming at the problem of cell position offset,the neighboring aggregation algorithm is used to calculate the average value of the neighboring area of each pixel of the recognition result to improve the accuracy of cell positioning.(4)For a data set with only cell location and category annotations but no regional annotations,a new cell ROI extraction algorithm for weak supervision was designed specifically,based on the existing point annotations,through the design of restrictive The Delaunay triangulation algorithm connects local cells of the same type and draws the region to obtain an ROI label similar to the artificial region label.Compared with the existing methods,Tri-NET captures the low-level texture features common to the cell recognition task,the cell location task and the cell ROI segmentation task at the same time for the shallow convolutional layer,so as to promote the accuracy of cell recognition together.This paper compares the two methods proposed in this paper with the recent end-toend cell recognition method on two datasets of different staining and different tissues.The experimental results show that,compared with other existing methods,MACECNN has good generalization in different stains and different tissues.The weighted F1 scores of different types of cell recognition tasks are improved by 2.13% compared with the existing methods.Compared with the MACE-CNN,Tri-NET,which uses ROI generation,has a similar F1 score,and the recognition process is not effected by background impurities.