Neuronal Morphology Classification Approach Based On Deep Learning Networks

The structure and function of brain are very complex.Human brain plan combines neuroscience and information science to research the structural characteristic and cognitive function of brain.Its goal is to use modern information tool to establish general standard of neuroinformatics database,and to retrieve,analyze,integrate and model neuroinformatics data at different levels.As the basic component and calculation unit of the nervous system,neuron has a variety of spatial geometry structure.The reasonable division of the type of neuron according to the geometry of neuron can better understand the characteristic and information transmission process of neuron,as well as the relationship between the structure and function of the brain.However,neuron has many kinds and characteristics,which make it very difficult to classify neuron accurately.Deep learning is a machine learning algorithm that can automatically extract abstract,nonlinear high-dimensional feature from complex data to solve the problem.Deep learning has made great progress in image understanding,speech recognition,natural language processing and other fields,providing a new computing method for complex pattern recognition problem.In this thesis,the deep learning network model is used to learn and extract the feature of different neuronal data,so as to better study the neuronal morphology classification of neuron.First of all,the neural classification data set in the study of neural geometric morphology classification is constructed.When the selected neurons are quite different,the difficulty of neural geometric morphology classification will be too low and the verification of generalization ability will be insufficient.In this thesis,the neuronal data sets of different types in C.elegans and zebrafish are constructed to verify the classification accuracy and generalization of the approach proposed in this thesis.These two kinds of neurons have high geometric similarity,so it is very challenging to classify them accurately.Secondly,in order to realize the effective classification of the geometry morphological neuron,the thesis first studies the fully connected deep neural network and the deep residual neural network,applies the skip-connection idea of the deep residual neural network to the fully connected deep neural network,then uses the idea of local cumulation to improve the deep residual neural network,and proposes a neuronal morphology classification approach based on locally cumulative connected deep neural network.This approach is applied to the geometric morphological classification of two kinds of neuronal data sets,and the influence of model parameter on the classification performance is analyzed.By comparing with several classical machine learning methods,the experimental result show that this method has higher classification accuracy.Finally,the traditional neuronal classification approach based on geometric morphology depends on the extraction and selection of neuronal spatial structure feature,which will lose a lot of useful neuronal classification information.Based on the geometry information of three-dimensional neuron,a neuronal morphology classification approach based on deep learning network is proposed.The characteristic of this approach is that it doesn't need to extract the quantitative geometric morphological feature of neuron.It can directly input the information of neuronal image into convolutional deep learning model to realize the classification and recognition of neuron.In this approach,the original neuronal data is reconstructed in three-dimensional voxel,and the adaptive projection process is used to form two-dimensional neuronal image data.A deep learning model based on a double convolutional gated recurrent neural network is constructed to classify neurons.Compared with the neuron classification approach based on feature extraction,the experimental result show that the approach has higher classification accuracy and good adaptability.