局部特征与点云配准引导下的神经元相似性度量方法  

Neuron Similarity Measurement Method Guided by Local Feature and Point Cloud Registration

作　　者：甄昊天 常令琛 祝继华[1] 朱恩涛 樊夏玥 李钟毓[1] Zhen Haotian;Chang Lingchen;Zhu Jihua;Zhu Entao;Fan Xiayue;Li Zhongyu(School of Software Engineering,Xi’an Jiaotong University,Xi’an 7100492;Center for Immunological and Metabolic Diseases,the First Affiliated Hospital of Xi’an Jiaotong University,Xi’an 710000)

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免疫代谢性疾病研究所,西安710000

出　　处：《计算机辅助设计与图形学学报》2024年第9期1480-1490,共11页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61902310)。

摘　　要：针对现有神经元形态学匹配与相似性度量方法难以有效处理大规模且结构复杂神经元数据的问题,提出一种由局部特征与点云配准引导下的神经元相似性度量方法.首先,利用全局特征进行大规模检索完成神经元相似数据的初筛;其次,基于深度卷积自编码器,对筛选后的神经元数据进行无监督的局部特征提取,实现两两神经元之间的粗配准,并借助迭代最近点算法,将具有空间树型结构的神经元匹配问题转换为点云的三维配准问题;最后,通过与全局特征下的检索结果进行融合,实现神经元形态数据的相似性度量.在NeuroMorpho公开数据集上抽取了19286个神经元,并与其他6种相似性度量方法进行对比实验,文中方法的Top-1和Top-50精度分别为0.981和0.721,均优于现有对比方法,验证了其在大规模数据集上的有效性与精确性.Aiming at the problem that the existing neuron morphological matching and similarity measurement methods are difficult to effectively handle large-scale and complex neuron data,a neuron similarity measurement method is proposed based on local features and point cloud registration.Firstly,global features are used for large-scale retrieval to complete the initial selection of neuron similar data.Secondly,unsupervised local features are extracted according to the selected neuron data,based on the deep convolutional autoencoder,in achieving the coarse registration between pair-wised neurons.Subsequently,the matching problem of neurons with spatial tree structure can be transformed into 3D point cloud registration problems based on iterative closest point algorithms.Finally,the similarity measurement of neuron morphological data can be achieved through the fusion be-tween the registration score and the retrieval results under global features.19286 neurons are extracted from the NeuroMorpho public data set and compared with 6 other similarity measurement methods.The accuracy of Top-1 and Top-50 of the proposed method are 0.981 and 0.721 respectively,both of which are better than the existing methods.Experimental results on large-scale datasets demonstrate the effectiveness and accu-racy of the proposed method.

关 键 词：神经元形态学 深度神经网络 相似性度量 点云配准 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]