基于三维深度神经网络的大规模神经元形态表征与检索方法  

Large-scale neuron morphological representation and retrieval based on a three-dimensional deep neural network

作　　者：常令琛 李钟毓[1] 樊夏玥 商增谊 景海婷 Lingchen CHANG;Zhongyu LI;Xiayue FAN;Zengyi SHANG;Haiting JING(School of Software Engineering,Xi’an Jiaotong University,Xi’an 710049,China;Med-X Institute,the First Affiliated Hospital of Xi’an Jiaotong University,Xi’an 710061,China)

机构地区：[1]西安交通大学软件学院,西安710049 [2]西安交通大学第一附属医院Med-X研究院,西安710061

出　　处：《中国科学：信息科学》2021年第12期2089-2101,共13页Scientia Sinica(Informationis)

基　　金：国家自然科学基金(批准号:61902310)资助项目。

摘　　要：高效准确的相似神经元检索方法是神经元形态分析的重要支撑.随着高精度显微成像、神经元示踪、人工智能等技术的发展,近些年出现了若干基于机器学习的神经元形态计算与分析方法,这些研究主要包括对传统神经元形态度量指标的统计分析,以及将神经元形态二维投影与深度学习结合的神经元量化表征方法,在神经元的特征提取、分类、相似检索等任务中均取得了不错的效果.不过随着越来越多的三维神经元数据被重建出来,以上方法都无法满足当前背景下对大规模神经元形态数据的细粒度表征、检索与分类需求.为此,本文提出了基于三维深度神经网络的大规模神经元形态表征与检索方法.首先,为了将神经元的三维空间拓扑结构转换成适用于深度神经网络的形式,我们设计了神经元空间形态的体素转换方法,将原始的神经元重构文件转换成三维体素的形式,极大地保留了神经元的三维空间拓扑结构.随后,考虑到当前神经元数据缺乏精细的分类标准,本文设计了基于三维卷积自动编码器的神经元形态表征算法,应用深度神经网络无监督地学习神经元体素数据的结构特点,得到神经元形态的量化表征,并以此设计端到端的相似神经元快速检索算法.最后通过实验验证本文所提出的方法,在9万余神经元数据中检索形态相似的神经元,实验结果显著优于其他基于神经元量化表征的检索方法.实验表明,本文方法可以更高效准确地检索相似神经元,为神经元的形态学分析、神经元单细胞分类等相关研究的关键问题提供支持.Efficient and accurate retrieval of similar neurons is an important support for morphological analytics of neurons.With the development of high-resolution microscopy imaging,neuron tracing,artificial intelligence,etc.,several methods for computing and analyzing neuron morphology based on machine learning have been proposed in recent years.These studies mainly include the statistical analysis of traditional measurement of 3D neuron morphology,as well as the quantitative representation of neurons by combining the two-dimensional projection of neuron morphology with deep learning.However,as more and more 3D neuron data being reconstructed,the above methods can no longer meet the current needs of fine-grained representation,retrieval,and classification of large-scale neuro-morphological datasets.Therefore,this paper develops a large-scale neuron morphology representation and retrieval method based on a 3D deep neural network.First of all,to transform the three-dimensional topology of neurons into a suitable format for the training of deep neural networks,we design a voxel transformation method for neuron spatial format,which can transform the original neuron reconstruction file into three-dimensional voxels.Then,because of the lack of detailed taxonomy for neuron data,we design a neuron morphology representation algorithm based on a 3D convolutional auto-encoder,which employs deep neural networks to learn the structural representation of neuron voxels in an unsupervised manner.Subsequently,given the learned neural quantitative representation of 3D morphology,this paper introduces an end-to-end fast search algorithm for the indexing of similar neurons.Finally,the proposed method is validated by experiments.Similar morphological neurons are retrieved from the dataset containing more than 90000 neurons.Experimental results show that our method achieves superior performance in comparison with other representational methods in the task of neuronal retrieval.The proposed method can retrieve similar neurons with more effici

关 键 词：神经元形态学 深度神经网络 三维体素 特征表达 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]