Adjacent Pixels Relation Based Cell Localization And Counting

The number and location distribution of the cells in pathological images reflect important information about some specifical diseases,and are useful for pathological analysis and disease diagnosis.Traditionally,cell localization,counting,and related disease diagnosis were performed by doctors with professional knowledge.Such labor intensive and time consuming works hinder the rapid diagnosis and accurate treatment.With the development of image processing technologies and their application in medical image analysis,cell localization and counting technologies have made a great progress.Detection and segmentation-based localization methods as well as regression-based counting methods achieve good results for images with simple cell morphological structure and of high contrast.However,their performances degrade significantly when dealing with complex clinical pathological images,where cells usually suffer from deformed shape,blurred edge,low contrast,and cell overlapping.Such degradation makes it hard for algorithm to meet the diagnosis requirement.In recent years,deep learning and its application research in image processing has made a great progress.For cell counting task,some break-through methods have been proposed by exploiting the outstanding feature extraction and expression capability of deep learning.Yet,in the case of pathological images with cells suffering from dense,overlapping,and blurred boundaries,it is hard to separate cell individuals based on cell detection and segmentation,resulting in missed detection and wrong localization.To design cell localization and counting system for pathology images with complex cells,this thesis proposes to leverage neighboring pixel relationship for accurate cell localization and counting.The detail is summarized as follow:First,by analyzing cell morphological structures and their distributed features,we find that the samples which are difficult to distinguish from each other are mainly located in some small dense areas.Accurate feature expression and extraction for cells in these areas is the key to high-performance algorithm design.In this thesis,we explore the characteristic of relationship between adjacent pixels in these regions,model the geometry position replationship between cell region pixel and its neighboring pixels,and encode them as unit direction field vector.Then,based on the analyzation of direction field vector image,we summarize its characteristics as follows: 1)Encoded directional field vectors for pixels in cell regions point to each corresponding nearest cell center;2)Direction field for adjacent pixels in different cell regions have opposite directions departing from each other.Such property is used to separate overlapped cells for localization and counting;3)In regions with blurred boundaries and low contrast,the coded vectors in cell regions have magnititude 1,and 0 for the background region,which helps for cell boundary delination.Finally,to deal with cells suffering from dense distribution,size variation,and overlapping between neighboring cells,we propose to generate the ground-truth direction field using geometry adaptive radius.Specificically,we dilate the cell centers with structuring element of adaptive size,giving rise to non-overlapped foreground cell region,on which we compute the unit direction vectors as the ground-truth direction field.This guides the convolutional neural network(CNN)model to predict the direction field for overlapping cells of varied size.Extensive experiments on four widely used datasets(i.e.,CRCHisto Phenotypes2016,Cell,MBM,and Mo Nu Seg)show that the adopted deep learning network can accurately predict the distribution of the direction field of pathological images.By leveraging the relationship between the direction field vectors of adjacent pixels,the accurate cell localization and counting can be achieved even in the case of cell deformation,occlusion,blurred boundaries,and low contrast.The proposed method outperforms some state-of-theart methods in both cell localization and counting.