Synapse Detection And Segmentation In Electron Microscopy Of The Complete Drosophila Brain

Synaptic connections,as a channel for transmitting signals between different neu-rons,play a very important role in the establishment of the function of the nervous system.Synapse size,shape,distribution,number,etc.are important information for understanding the function and plasticity of the nervous system.So far,only the electron microscope can observe the synaptic-level connection structure with nano-scale resolu-tion.A nano-scale electron microscope scan image volume of the complete drosophila brain contains 40 trillion pixels.Faced with such a large-scale image data,how to ef-ficiently extract effective synaptic connections has become an urgent problem to be solved.The main difficulty of this work is the complexity of the background content of the electron microscope image and the difficulty of labeling.In addition,due to the anisotropy of whole Drosophila brain electron microscope data and the existence of ax-ons corresponding to multiple dendrites in the insect brain,many existing methods are not suitable for Drosophila data.This thesis focuses on the task of detecting and seg-menting synaptic connections in whole Drosophila brain electron microscope data,and conducts research from the following two aspects:(1)Design a synaptic cleft detection and segmentation algorithm using a two-stage method based on a 3D convolutional neural network.However,due to the complex background of the electron microscope image,using the deep neural network to directly segment the synaptic cleft will cause some subcellular structures to be unable to be correctly distinguished.In addition,the 3D information of the electron microscope image is also particularly important for detecting the synaptic connection.Based on this consideration,this thesis uses a 3D convolution network to make full use of the 3D information of the electron microscope image,and then we adopts a two-stage method to initially detect the area of the synaptic cleft and then perform detailed segmentation in the local area.(2)Proposes the idea of segmenting presynaptic axon to help obtain synaptic con-nections.Compared with the synaptic cleft,the presynaptic axon is larger and the char-acteristics are more obvious.By detecting and segmenting the presynaptic axon,we can indirectly observe the distribution of synaptic connections.In order to train the presynaptic axon segmentation neural network,this thesis studies two presynaptic axon labeling methods,three-dimensional sparse labeling and two-dimensional dense frame labeling.The effectiveness of the two labeling methods is compared and analyzed through experiments.In order to reduce the dependence on labeled data,this thesis explores a semi-supervised method based on consistency regularization to efficiently use unlabeled data.Finally,the presynaptic axon is segmented in the whole brain to lay the foundation for subsequent biological analysis.In summary,the work of this thesis focuses on research hotspots,and provides powerful technical support for the increasingly vigorous brain science analysis.We provides two ideas for obtaining synaptic connections.The detection and segmenta-tion of synaptic clefts combine 3D convolution and two-stage methods to directly ob-tain synaptic connections,while presynaptic axon detection and segmentation provide a new idea for detecting synaptic connections,focusing on presynaptic axon with obvious characteristics.In addition,the two presynaptic axon datasets we labeled also promote the progress of related research work.