Feature Extraction And Classification Of Bone Mesenchymal Stem Cells Based On Machine Learning

Bone mesenchymal stem cells(BMSCs)are stem cells which support hematopoiesis,promote hematopoietic progenitor cell proliferation,regulate bone marrow microenvironment and have multipotential differentiation potential.BMSCs also play an important role in tissue immune regulation and can be utilized as seed cells in tissue engineering for disease treatments.Currently,there is no impeccable evaluation system for the physiological function of BMSCs,which mainly relies on the clinical experience of doctors and the determination of protein expression.Cell image segmentation and recognition based on morphology is the key technology for evaluating physiological function objectively.The main works of this paper include:After cell culture,BMSCs images are obtained by fluorescence confocal laser scanning microscopy,and the images are optical sectionings with different staining information of nuclear matter.For image degradation of optical sectionings,blind deconvolution is utilized to reduce the degradation effect of defocusing surface,so as to complete image restoration.Then,according to the reconstructed images,the weighted average method and the OSTU method are used for gray processing and binarization.Based on the binarized image,the Moving-Cube algorithm is used to reconstruct the three-dimensional cell contour to obtain the three-dimensional structure of cell.On account of limitations of biological experimental operation and the shooting angle of confocal microscopy,the three-dimensional structure reconstructed by BMSCs optical sectionings has various attitudes and positions,which affect the accuracy,consistency and stability of cell feature acquisition and classification.Therefore,a method is proposed to obtaine spatial long axis in the three-dimensional contour,and a rotation method of coordinate axes is proposed to re-slice based on spatial long axis for obtaining the maximum cross section as the datum plane of feature extraction.Morphological characteristics of the cell image are calculated based on the datum plane,and the genetic algorithm is used for feature analysis and selection.Based on the selected features,support vector machine with different kernel functions and multilayer perceptron with different network structures are used for classification and identification,verified the validity of the maximum cross-section algorithm.The recognition accuracy is up to 97.8% compared with the performance of different structure classifiers.Finally,convolution neural network is used to input the maximum cross-section image of the cell,with data augmentation to expand the sample.Combined with tansfer learning,the Inception V3 is used to the classification and recognition of BMSCs successfully.Compared with the traditional machine learning methods,this method has best performance and the accuracy is up to 98.9%.In this paper,to obtain the feature vector,the pretreatment and 3D reconstruction methods based on the optical sectings of BMSCs,the method to calculate the spatial long axis based on the three-dimensional structure of the cell,and the method of obtaining the maximum cross section by the coordinate system rotation are proposed to establish a solid foundation for obtaining feature vectors and classification recognition.The classifiers have excellent performance and provide basis for objective evaluation of BMSCs function.