Substructure Analysis And Retrieval Of Three-dimensional Reconstruction Of Neuron

As basic unit of the brain,neurons have a wide variety of forms and complex shapes.Tracing the neuron from the microscopic image and reconstructing its digital model is a key way to understand the neuron.Because tracing neuron manually takes time and effort and the existing auto-tracing algorithms lack good applicability,it is of great significance to study the method of reconstructing neuron from three-dimensional images rapidly and to use the reconstructed data for neuronal morphology analysis.This research explory substructures on three-dimensional neuron reconstructions.Via obtaining the accurate substructures and retrieving similar substructures on reconstructions,it is possible to help researchers obtain neuronal digital models efficiently and find a way to analyze the performance of neuron tracing algorithms.The main researches are as follows:(1)Exploring substructures with high credibility on neuron reconstruction.For reconstructions generated by dozens of different auto-tracing algorithms of a same neuron,the local alignment algorithm is used to find out the topological relationships between any two of them,and then three strategies are designed to find overlapping nodes.Finally,the reconstruction motif of this neuron can be obtained through connecting,pruning,and deleting outliers.Reconstruction motifs generated by different parameters under the three strategies are compared in the experiment,one of the strategies with great applicability is choosen for further analysis on four neuron tracing algorithms: APP2,NeuroGPS-Tree,Neutube,and Snake.The results show that the reconstruction motif can represent the accurate parts of the automatic reconstruction,which can be used as the initial stage for manually reconstructing.At the same time,the accuracy and reappearance of the motif can also reflect the quality of the reconstruction.(2)Proposing a framework for retrieving similar substructures on neuron reconstruction.Given a reconstruction and an interest border,the framework takes advantage of the characteristics of tree structure of neuron to extract the substructure within the border of interest as the query,and then generates numerous candidate substructures on the entire reconstruction.Next,the retrieval framework implements the global search function to calculate the 19 morphological features of the query and the candidates and then sort them.Finally,multiple substructures which are most similar to the query among the candidates are obtained.Multiple sets of visual comparison experiments show that the framework has good applicability and the retrieved substructures have similar nodes,branch numbers,and curvatures as the query.(3)An improved method of constructing substructures.For a neuron reconstruction without connection characteristics,the spatial segmentation method is proposed based on similar substructure retrieval framework.This method divides the entire neuronal space into multiple ordered subspaces and establishes the map between each node on the reconstruction and the subspace.The query and candidates can be quickly constructed.Experimental results show that the combination of spatial segmentation method and retrieval framework can achieve similar retrieval of substructures on the reconstruction which consists of multiple disconnected segments.The exploration of neuron reconstruction motif and the implementation of substructure retrieval functions facilitate the analysis of the use of reconstructed data by researchers.At the same time,it promotes the classification and function research of neurons based on morphology.