Identification Of Single-neuron Axonal Boutons Integrating Density Peak Clustering And Convolutional Neural Network

The brain is the central structure of human nervous system,and mapping brain neural circuits is a core goal of modern brain research.Studies of brain neural circuits have important guiding significance for treatment of brain diseases and development of artificial intelligence.The neural circuits are composed of massive neurons that establish synaptic connections with each other.Therefore,synaptic connectivity inference is the key to analyzing neural circuits.Single-neuron synapse identification helps neuroscience researchers to obtain specific sites and number of synapses,and then to quantify and analyze the information transmission patterns between neurons.At present,synapses are mainly identified through electron microscopy images,but it is difficult for electron microscopy to image the complete shape of a single neuron in the centimeter range,and only local synapses can be obtained.Submicron-resolution brain-wide neuron morphological data from light microscopy provides the possibility to obtain complete synapses of a single neuron.However,current neuronal morphology reconstruction researches in light microscopy images mainly focus on the skeletal tracing of neurites,and the identification of axonal boutons is still lacking(axonal synapses appear as axonal boutons in light microscopy images).Few existing axonal bouton recognition studies are limited by their method principles and can only be used for local axonal images with limited recognition accuracy generally.In response to the above needs and current situation,this thesis develops a method system for single-neuron axonal bouton identification and uses this method to conduct a preliminary study on the spatial distribution of single-neuron axonal boutons.(1)I propose a recognition method for neuronal point-like structures based on density peak clustering.In this thesis,by designing a local search strategy and defining the local density of scale correlation,a method suitable for identifying point-like structures in large-volume three-dimensional(3D)neuronal images is developed.Soma recognition experiments verify the effectiveness of the density peak clustering recognition method for identifying point structures in 3D neuronal images.Through the identification of simulated point structure data,the inherent characteristics of the proposed method are analyzed,including the effectiveness of segmenting the touching point structures,the robustness to scale parameters and the approximately linear time complexity.Through experiments of axonal bouton identification in local images,it is verified that the density peak clustering recognition method can be used for the preliminary identification of axonal boutons.(2)I develop an automated method for recognizing single-neuron axonal boutons integrating density peak clustering and convolutional neural networks.First,the morphological characteristics of neuronal axonal boutons and the challenges of the previous recognition method based on density peak clustering in identifying axonal boutons are analyzed.Then combined with these morphological features,an automated method for single-neuron axonal bouton identification based on a two-step recognition strategy is established: initial recognition of underlying axonal boutons based on improved density peak clustering algorithm and secondary identification of initially detected axonal boutons based on convolutional neural network model.A convolutional network architecture is designed according to the morphological characteristics of the axonal boutons.The network can effectively characterize the morphological characteristics of the axonal boutons and help distinguish axonal boutons from non-bouton axonal swellings.The introduction of convolutional networks solves the problem that the density peak clustering recognition method is difficult to accurately describe the morphology of axonal boutons.This method has achieved accuracy and recall of about 0.9 on various datasets,which proves the effectiveness of my method in identifying single-neuron axonal boutons.Furthermore,a two-color imaging experiment is designed to verify the high correlation between the automatically identified axonal boutons and the protein-specific axonal synapses.About 80% of the axonal boutons can find the corresponding protein-specific axonal synapses.(3)I conduct a preliminary study of the spatial distribution of single-neuron axonal boutons using the developed recognition method above.The spatial distribution of multiple neuron axonal boutons is analyzed from multiple aspects,including the distribution of single-neuron axonal boutons in different brain regions,the distribution of single-neuron axonal boutons of subtypes and the spatial line density of single-neuron axonal boutons.Studies show that the number and the interval of axonal boutons in different brain areas are very different;5 subtypes are summarized according to the distribution pattern of axonal boutons in different brain areas;there is no strict linear relationship between neuronal projection fiber length and axonal bouton number.The single-neuron axonal bouton identification method established in this thesis advances neuron reconstruction research from preliminary skeletal tracing to synapse-level single-neuron fine morphological reconstruction,thereby helping neuroscience researchers to more accurately obtain the transmission nodes of neuronal information and more fully analyze the information transmission patterns of neural circuits,cell type,synaptic plasticity,etc.The axonal bouton analysis tool established in this thesis has been initially applied by biological colleagues in my research group.