Research On Measure And Classification Of Cell Dynamic Morphology In Microscopic Videos

Cell morphology is related to the cell physiology,and has been used as a proxy measurement of cell status in biomedical studies.In recent years,thanks to the advances in medical image processing and computer technology,researchers are able to observe the cells with higher fidelity and analyze the cell morphology more efficiently.Consequently,cell deformation has become a new target in cell biology research.Cell deformation provides more meaningful clues of the cell physiology,and cell status can be identified according to cell dynamic morphology in a short period of time.However,how to characterize and analyze the dynamic morphology remains a challenge task.Therefore,the characterization and classification of cell dynamic morphology is significant in cell biology research.In this study,the dynamic morphology of T-lymphocytes are observed and recorded to build the microscopic video database of cell dynamic morphology.The algorithms based on image features are used to interperate the dynamic morphology,and temoral features are designed according to the deformation reguliarities to characterize and classify the dynamic morpholohy.In addition,deep learning algorithms are introduced to automatically undersrand and classify cell dynamic morphology.The main contents and contributions of this study are as follows:1.The construction of a cell dynamic morphology database and preprocessing framework.By collaboration with Beijing Youan Hospital,lymphocytes were collected from the blood of mice,which underwent skin grafts.Then the dynamic morphology of the lymphocytes was recorded as videos,and quality control and annotation were performed on the data.According to the properties of the data,a preprocessing framework is developed to normalize cell scales and align cell poses.Through the collection and preprocessing,the database is built to be employed in the following studies.2.A scheme of quaternion generic Fourier descriptor(QGFD)for color object recognition is developed.Based on the study of image features,QGFD is developed and achieves remarkable performance on color object recognition.Then QGFD and other image features are applied to cell dynamic morphology characterization and classification.The advantages and problems of the feautres are demonstrated,and the challenges in the interportation of cell dynamic morphology are presented.Powerful characterization algorithms of cell dynimac morphology are developed aiming to the challenges.3.The study of temporal feature extraction and analysis algorithms of cell dynamic morphology.Three feature extraction algorithms are developed to consider the temporal regularity and spatial heterogeneity of cell dynamic morphology.The algorithm based on the run length of image match is first introducted to characterize the dynamic morphology without image feature selection.The similarities between cell morphologies are quantified by the energy function of image match,and the temporal regularities of the dynamic morphology are captured by the run length.Subsequently,based on the observation of the video data and the inference of the image features,an algorithm is presented to characterize the temporal regularities of cell local deformation.The temporal regularity and spatial heterogeneity of the dynamic morphology are comprehensively considered by the local temporal features.Finally,local deformation patterns are identified using the local temporal features and an unsuperived clustering algorithm,to simplify the features and boost the classification.The experitmental results deomstrate the ability of the proposed algorthims,and the advantage of considing the temporal regularity and spatial heterogeneity of cell dynamic morphology is validated.4.The study of cell dynamic morphology classification using deep learning.Inspired by the advantages of deep learning,convolutional neural networks are applied to cell dynamic morphology classification by converting the data from videos to images.Cell contour spectrogram is generated to convert the dynamic morphology in videos into 2D images.And data augmentation methods of the generated images are descripted,which increase the size of the training set.Then,to fully understand the portential of deep learning,three scenarios are respectively performed to install the convolutional neural networks on the classification of cell dynamic morphology.According to the performances in experiments,the convolutional neural networks are successfully applied to the classification of cell dynamic morphology,by converting the data from videos to images.Cell dynamic morpholohy is effectively inteperated and classified by the convolutional neural networks,and the installation scenarios of treanfer learning present remarkable performances.