Research Of Key Technologies For Microscopic Cell Image Sequence Morphology Analysis

The quantitative analysis of morphology and motility for target cells in microscopic cell image sequence,it is very meaningful for understanding and analyzing the biological process of life.The morphological changes of cells are often associated with certain physiological and pathological process.There is an extensive application value that analyzing the cell's dynamic behavior to reveal the relationship between the organism's physiological or pathological status and cell morphology change.We take the pluripotent stem cells image sequence as the research object,and made a research on key technologies for cell image sequence morphology analysis.The research topic is mainly focus on cell image sequence recovery,contour boundary point matching,extracting deformation factor and morphological change pattern classification,including the following four aspects.(1)Research of cell image sequence restoration method based spatial-temporal Gaussian curvature regularizationWe propose an image sequence restore method for continuous cell image sequence denoising and deblurring.With the correlation of spatial and temporal domain,our method firstly taken the cell image sequence as a spatial-temporal volume,and introduced the spatial-temporal Gaussian curvature regularization to build the restoration model.The regularization model can effectively enhance the smoothness of the model's optimal solution;therefore the recovery one is more close to original image.Then,during the solution of the optimum model,we combined the augmented Lagrangian multipliers and division algorithm to find solution of optimization problem iteratively.Lastly,to verify the effectiveness of our method,we do denoising and deblurring experiment on six group different cell image sequence data set,respectively.Experimental results show that comparing with the other image sequence restore method based on spatial-temporal volume,our method get a higher quality of image restoration effect and more natural cell image.(2)Research of cell image sequence contour points corresponding method based adjacent tensor matchingThe corresponding between the contour points of cells image sequence is the premise condition of deformation factor extraction,and it is usually solved by graph matching method.We proposed an improved high-order constraint graphics matching method to implement the one-to-one correspondence between the contour points of cells image sequence.Firstly,the graphics is coded by the adjacent tensor,compared with the existing hyper graph matching algorithm,the coding data storage memory gained optimization essentially.Then,the triple descriptor was used to express the graph structure of cell contour point sets,the graph was expressed as the three dimensional tensor,and then convert them into matrix form that easier tackling.We solved the matching problem by graduated nonconvexity and concavity optimization based on gradient optimization method.Lastly,the comparison experiment was conducted on synthesis point sets and actual cell image sequence contour point sets by different homologous point set matching method,respectively,the experimental results verify the effectiveness of our method.(3)Research of cell image sequence deformation model based structured matrix decompositionA high dimensional shape space is formed by the continuous cell image sequence contour,and the deformation between these shapes can be expressed as a linear combination of the low dimensional sub shape space.In order to meet the demand of quantitative description and analysis of cell image sequence deformation,Firstly,we proposed a deformation factor of shape sequence based on graphics regularization and structured matrix decomposition.Then,the optimization solution of deformation factor is solved by adjacent gradient descent,and made decomposition.With good smoothness,sparse and local characteristics,the deformation factor can viewed as the parameters what used to quantify the characterization of cell shape and deformation in image sequence,and it is a good way to depict the dynamic deformation process of shape sequence.Lastly,this method was applied to a synthesis shape deformation sequence and a real cell image sequence;the experimental results verify the effectiveness of deformation feature extraction and analysis of shape sequence.So this research laid the foundation for subsequent cell image sequence classification based on the morphological changes.(4)Research of cell image sequence classification method based linear chain condition random fieldThe study realized cell image sequence classification by probability model based on linear chain condition random field.The multi-pattern classification problem of cell pathological changes can be expressed as probability problem that what kind of cell image sequence is it like.Firstly,the cell image sequence classification problem was modeled as a multi-class classifier based on linear chain condition random field,and this was a conditional probability distribution model with respect to classes.Then,the model parameter was estimated by discrimination learning algorithm based on margin maximization criterion between different classes.Lastly,we finished the image sequence classification of cell morphological changes according to the input feature vectors,which included deformation factor and dynamic texture describing the internal motion for cells image sequence.The result of numerical classification experiment and actual cell image sequence classification shows that our method has reached higher classification accuracy with good adaptability and stability.