Study On Motion Correction And Segmentation Of Biomedical Microscopy Cell Images

Microscopy cell image processing is an important research direction in imagerecognition field. Microscopy cell image processing combines the knowledge fromseveral directions such as image processing, pattern recognition and computer vision,and it is broadly applied in many fields. With the rapid development of science andtechnology, the development in the biomedical field has been increasinglydemanded by human. To explore the pathogenesis of the diseases and to reveal theorigin and formation of life, by quantitative research in molecular level, is the majordirections for biomedical field in current days and also in future. The biomedicalfield evolves, starting from qualitative analysis to high-resolution structuralquantitative evaluation, giving a demand for quantitative processing of electricalmicroscopy images. As a result, the research for cellular resolution image processingis significantly meaningful for biomedical field.This paper is focused on two specific problems for methodology research anddevelopment, which are motion correction in two-photon fluorescence microscopyimaging of awake head-restrained mice and segmentation of differential interferencecontrast cell microscopy images, respectively. Motion correction is a specificapplication of image registration in temporal image sequences. Image registrationand image segmentation are both significant fields in image processing and theyrelates to each other intensively.In motion correction of two-photon laser scanning microscopy images ofawake mice, the traditional hidden Markov model often fails and gives the wrongmotion estimation during the running stage. To solve this problem, we propose aspeed embedded hidden Markov model. This model firstly uses an exhaustivesearching method to obtain a preliminary offset estimation result for every line ofimage sequences, and then embeds this preliminary speed estimation result into thestate transition probability for speed compensation. This model overcomes theinherent drawback of the traditional hidden Markov model which assumes that thehighest probability is assigned to the case with no motion. Based on the simulateddata and real data, experimental results validate the speed embedded hidden Markovmodel can get more accurate motion correction results than the traditional hiddenMarkov model. As for the reference slice selection problem in initialization of motioncorrection of two-photon laser scanning microscopy images of awake mice, thetraditional intensity based reference slice selection method often ignores the changesof spatial geometry information between time serial slices, resulting in the lowaccuracy of motion correction. To solve this problem, we propose a novel bestreference slice selection algorithm. This algorithm no longer uses the pixel intensitydifference as registration metric, but employs the combination of mutualinformation and mean absolute error. This algorithm takes into account the spatialinformation given by the mutual information, which avoids the local maxima andmisalignments that mean absolute error would brings if used only. Based on thesimulated data and real data, experimental results validate the best reference sliceselection algorithm can obtain more accurate motion correction results than thetraditional typical algorithms.For edge detection problem of differential interference contrast cell microscopyimages, the traditional intensity based cell edge detection algorithms suffer fromsome imaging drawbacks such as uneven illumination and low contrast, thus theresults of edge detection is not promising. To solve this problem, a phasecongruency based edge detection method for cell images is proposed. This methodmodifies the phase deviation weighting function within phase congruency modeland improves the sensitivity of phase congruency model when capturing the localstructural significance. The edge detection results of neural stem cell images and redblood cell images indicate that the phase congruency based edge detection methodfor cell images brings more accurate edge information for cell boundarysegmentation.In segmentation problem of differential interference contrast red blood cellmicroscopy images, the inherent drawbacks of red blood cell images results in thedifficulties for traditional active contour model algorithm to take advantage of theregional statistical information and edge feature information, and off-centerreference point leads to the unsatisfactory cell segmentation results for thegeneralized version of subjective surfaces. To solve these problems, a subjectivesurfaces based cell image segmentation method is proposed. This method introducesthe edge detector which is based on local phase information and proposes a newvariation scheme for stretching factor within the framework of subjective surfaces.Experimental results indicate that the subjective surfaces based cell imagesegmentation method efficiently improves the segmentation accuracy of cell imagesand can still work well relatively even if the reference point is located nearby the cell boundaries.