Large-scale Neuronal Population Reconstruction And Neuronal Morphological Analysis

Morphological reconstruction and analysis of large-scale neuronal populations are important topics in brain science and have been extensively studied in computational neuroscience.Neuron morphology is crucial for neuronal activity,plasticity,and connectivity,which are considered to be well correlated with the neuron’s physiological properties and functionalities.With the development of optical microscopy imaging,morphological reconstruction of neuronal populations from ultra-scale optical microscopy brain images and the analysis of these neuronal morphological data becomes more and more important to investigate the mechanism of the nervous system,analyze brain developments,and facilitate research of brain diseases such as dementia and Alzheimer’s disease.Many efforts have been devoted to neuron reconstruction and neuronal morphology representation with hand-crafted features.With the rapid development of deep learning,a series of methods based on deep neural networks have been proposed for neuron reconstruction and neuronal morphological analysis.However,the low image quality,complex neuronal morphologies,deficiency of data annotations,and the sheer volume of optical microscopy images together pose significant challenges in neuron reconstruction and morphological analysis.In this dissertation,we mainly study semisupervised and unsupervised techniques for neuron reconstruction and morphological analysis,in order to achieve fine-grained morphological observation of several neurons,then large-scale neuronal population reconstruction,and finally efficient and accurate morphological analysis of thousands of neurons.This dissertation first proposes a progressive learning based neuron reconstruction method to reconstruct neurons from optical microscopy images,which enables morphological observation of individual to several neurons.Then,this dissertation designs a large-scale neuronal population reconstruction framework to reconstruct neuronal populations from ultra-scale optical microscopy brain images,which enables morphological observation of hundreds to thousands of neurons.At last,this dissertation proposes a deep hashing based morphology-aware contrastive graph neural network to the largescale neuron morphological analysis,which enables efficient and accurate morphological analysis of tens of thousands to hundreds of thousands of neurons.The main contents and contributions are summarized as follow:(1)Neuron Reconstruction from Optical Microscopy ImagesGiven optical microscopy images with low signal-to-noise ratio and noncontinuous intensities of neurite segments as input,existing methods that rely on handcrafted features usually fail to reconstruct complete neurons.Though deep learning techniques can improve the performance of neuron reconstruction,they rely on extensive manual annotations of neuronal voxels for network training.To alleviate the heavy burden of voxel-level annotation,a progressive learning-based neuron reconstruction method is proposed in this dissertation.It leverages the advantages of traditional neuron tracing techniques and deep segmentation networks.While conventional neuron tracing techniques do not require expensive manual labels,they can be used to produce pseudo-labels of neuronal voxels.On the other hand,a deep segmentation network is expected to learn more comprehensive neuronal features from noisy labels and extract more complete neuronal signals from noisy optical microscopy images.Based on a progressive learning scheme,the neuron tracing module and the deep segmentation network can mutually complement and promote each other to improve neuron reconstruction progressively.(2)Large-scale Neuronal Population ReconstructionExisting methods in dealing with large-scale neuronal population reconstruction are faced with many problems,including complex and diverse neuronal morphologies,huge memory requirements,and high computational costs.This dissertation proposes a new large-scale neuronal population reconstruction framework,which can automatically reconstruct dense neuronal populations from ultra-scale optical microscopy images.In this framework,an effective block propagation strategy is designed to trace neuronal populations across image blocks in an adaptive order and reconstruct largescale neuronal populations under limited resources.To reconstruct structurally continuous neuron structures across image blocks,a neurite fusion algorithm is proposed to fuse overlapped neurites in adjacent blocks seamlessly and smoothly.This method introduces the progressive learning-based neuron reconstruction algorithm to reconstruct more complete neuronal populations without relying on expensive manual labels.(3)Large-scale Neuronal Morphological AnalysisThe heterogeneity of morphologies,absence of annotations,and the sheer volume of neuronal morphological data pose significant challenges in neuronal morphological analysis.Many studies have been conducted to quantitatively describe neuronal morphologies using predefined measurements with limited representation ability,which are usually inadequate for distinguishing the fine-grained differences among massive neurons.To address these problems,this dissertation proposes a new neuronal morphological analysis method based on graph neural networks.A morphology-aware graph neural network is proposed to learn neuronal morphological representations efficiently.In addition,a contrastive learning mechanism with new neuron augmentation operations is introduced for unsupervised neuronal morphological representation learning.Subsequently,for fast retrieval in large-scale neuron datasets,we further design a deep hashing graph neural network.By introducing an improved deep hash algorithm,this method can be trained end-to-end to learn binary hash representations of neurons,which significantly improves retrieval efficiency.