基于卷积自编码器与点云配准的神经元形态相似性度量方法

Similarity Measurement of Neuronal Morphology Based on Convolutional Auto-Encoder and Spatial Registration

神经元形态与其功能密切相关,随着神经元示踪技术的进步,越来越多高质量的三维神经元形态学数据得以重建。针对当前体量越发庞大的神经元形态学数据,提出一种基于深度卷积自编码器和空间配准的神经元形态相似性度量方法,通过快速对比与精细对比两步完成从整体到分支的神经元形态分析度量框架,实现高效、精确的对比算法。实验所用的99453个三维神经元形态数据来自NeuroMorpho数据集。实验中,相比于现有的精细对比算法,该算法的实现速度加快了20倍,同时可应用于任意神经元形态学数据,无需提供其他的先验条件,通用性强。对于已在脑图谱模板中配准的神经元,选取233个uPNs神经元作为验证数据,可实现97.39%的检索精度;对于未配准的神经元,选取3种类型的神经元数据进行验证,包括:495个谷氨酸能神经元、389个DA神经元、249个锥体神经元,可分别达到91.7%、93.79%、83.1%的检索精度。所提出方法可为神经元类型鉴定、将神经元形态与特性进行关联分析提供支持。

Morphology of neurons is closely related to their function.With the advancement of neuron tracing technology,more and more high-quality digitally reconstructed 3D neuron morphology data are emerging.A twostep neuron morphology measurement framework,based on deep learning and 3D spatial data registration technology,was proposed for the computational analysis of 3D neuronal structures.Through fast comparison and fine comparison,the framework could be used for the growing volume of neuron morphological data from the whole neuron to the branch.99453 neurons from the NeuroMorpho dataset were selected for the experiment.Compared with the existing fine comparison algorithms,this framework was more than 20 times faster with good universality,which could be applied to any neuron morphological data without other prior conditions.For the neurons registered in the brain atlas template,233 uPNs were selected as the validation data,and 97.39%retrieval accuracy was achieved.For unregistered neurons,three types of neuron data were selected for verification,including 495 glutamatergic neurons,389 multi-dendritic-dendritic arborization neurons,and 249 pyramidal neurons.The retrieval accuracy reached 91.7%,93.79%and 83.1%respectively.Our proposed method is expected to be used for neuron type identification and correlation analysis of neuron morphology and characteristics.