The Study Of 3D Reconstruction Of Synapse Based On Serial Section Images

Brain science is a subject to research not only the cognition of human,animals and even machines but also the nature and regularity of intelligence.It is the core of researching brain science to understand the structure and material base of brain function.Researching of brain map is the key to explore the brain how to work,reveal cerebral diseases how to produce,and develop the brain computing.It is also the strategic high ground of brain science researching,and brain science is extremely possible to make a great breakthrough in the field.It is the foundation of drawing the brain map to reconstruct microstructure of brain.Neuro system is a complex network structure composed of large amount of connected neurons.To explore and understand the function of neuro system,the structure mapping information with high quality is necessary.Messages in brain map vary with the different scale.Formally,there are three different scales to research:macro scale in millimeter resolution,meso scale in micrometer resolution and micro scale in nanometer resolution.Nowadays,the method of auto-analyzing and ability of reconstruction in synaptic level are relatively lagged behind,reconstructing massive data is mainly relies on manmade working,which is not only inefficiency,but also error prone.The 3D reconstruction of large amount of neuro structure has been introduced bottlenecks,it has become a top priority to propose effective and automatic approaches to increase the speed of researching.This thesis relies mainly on the micro-scale reconstruction platform made by research group for microscopic reconstruction and analysis in Institute of Automation Chinese Academy of Sciences(CASIA).The main work of the thesis is to accomplish the 3D reconstruction of synapse in registered serial section EM images,by proposing a serial of automatic methods to detect and segment synapses.The method is composed of three parts.Firstly,synapses are detected based on machine learning methods.In the stage,Ada Boost algorithm are used to detect synapses roughly above all,where recall is important while precision is reasonable.After that,context cues and random forest methods are used to upgrade the precision in sequence.Secondly,detected synapses are segmented with mathematical morphological operation and region growing methods in good form structure.While segmenting,thresholding,erode and morphological open operations are used first of all.With the primary segmentation results,seed regions are gained by means of fitting and Dijkstra.According to the definition of region growing,points around seed region are appended with certain rules,and synapses are segmented well at this point.Finally,3D synapses are reconstructed by overlaying synapses segmented in registered serial section images in sequence.Approach proposed here can be applied to both isotropic and massive anisotropic serial section images to reconstruct synapse.