CASIA OpenIR  > 类脑智能研究中心  > 神经计算及脑机交互
基于深度学习的微观脑连接图谱重建方法研究
肖驰
Subtype博士
Thesis Advisor韩华 研究员 ; 谢启伟 教授
2019-05-24
Degree Grantor中国科学院大学
Place of Conferral北京
Degree Discipline模式识别与智能系统
Keyword微观脑连接图谱 电子显微图像 卷积神经网络 深度学习 图像分割
Abstract

脑图谱的微观重建是探索和理解脑神经功能的重要环节。利用电子显微镜对生物脑组织进行成像,然后对神经元和超微精细结构进行识别和重构,得到脑神经环路微观结构的工作是类脑智能和大脑逆向工程的重点方向。对于神经环路结构解析、神经性疾病机理探索、神经功能和人工智能等研究具有重要的意义。然而,微观脑图谱的重建中存在电镜图像背景复杂、精细结构形态差异大等问题,直接应用现有基于深度学习的识别和分割方法难以取得满意的效果。当前国际上主要的微观脑图谱重建工作大量依赖于人工标记与校验,实现大规模的神经环路重建往往需要耗费上万人工时。

 

本论文主要研究微观脑图谱中突触、线粒体和神经元的重建算法。针对现有方法的不足之处,分别设计适合目标形态特点的深度学习网络模型,突破精细结构自动识别、神经元密集分割等技术瓶颈,在提高算法精度和效率的同时减少对人工的依赖。论文的主要成果和贡献如下:

 

  1. 针对突触重建算法在复杂背景下识别率低、效率低等问题,本文提出了一套高精度的Coarse-to-fine突触自动重建方法。该算法首先通过基于深度学习的检测网络获取突触的具体位置,随后根据突触的连续性信息筛除假阳性结果,最后采用最短路径和GrabCut算法获取突触间隙的精细分割与重建结果。在ATUM-SEM和FIB-SEM数据中,该方法的检测准确率(92.8%和93.5%)和分割准确率(88.6%和93.0%)超过了目前大部分的突触识别算法,在计算效率上也达到了较高水准。此外,本文通过该方法完成了大体量电镜数据中突触的3D重建和形态参数统计等工作,为神经学家研究突触提供了可靠的分析数据。

 

  1. 针对线粒体重建中的目标形态差异大、图像噪声等问题,本文提出了基于3D全卷积深度监督网络的线粒体重建方法。该方法通过3D卷积网络能够有效利用电镜数据的三维空间信息,网络中U-Net配合残差卷积模块的结构可以充分融合不同尺度的特征,更好地识别不同形态大小的线粒体,深度监督的策略有助于避免训练中的梯度消失问题。该方法在多个线粒体数据集中得到了同期最好的结果。在此基础上,本文获取了大体量电镜数据中线粒体的3D重建结果和形态学等参数,并重建了小鼠运动皮层胞体和树突中的线粒体,证明树突中线粒体依靠“线粒体管道”连接在一起,为研究神经元中线粒体功能提供了新思路。

 

  1. 针对神经元重建中的图像背景复杂、神经元结构差异明显等问题,本文提出了基于上下文残差网络的神经元重建算法。该方法可以结合不同尺度的感受野信息,有效区分神经元细胞膜与其他细胞器,网络中的亚像素卷积层能够减少上采样过程中细胞膜信息的损失,后续采用的Lifted Multicut等算法则用于获取神经元的立体分割结果。该方法在ISBI神经元分割国际挑战赛中排名2/171,并在鼠脑梨状皮层数据集中获得了同期最好的结果。基于该套神经元重建算法,本文密集重建了体量为33*30*40 um3大小(单幅图像尺寸超过10000*10000,序列数量超过1000片)的鼠脑皮层区域,在少量人工校验的情况下获取了其中大部分的神经元结构。
Other Abstract

Microscopic connectome reconstruction is an important step in exploring and understanding the neural function of brain. The reconstruction work consists of imaging the brain tissue with electron microscope, identifying and reconstructing the neurons and ultrastructure, and obtaining the microstructure of the brain neural circuit, which is of great significance for the research of neural circuit, neurogenic diseases, neural function and artificial intelligence. However, the reconstruction of microscopic connectome is a tough task due to several challenges, e.g., the complex background of EM (electron microscopy) image, the various sizes and shapes of the ultrastructure, etc. It is difficult to obtain satisfactory results by applying the existing deep learning-based methods. Therefore, the current reconstruction work relies heavily on manual label work, it takes thousands of human work hours to achieve large-scale microscopic connectome reconstruction.

 

In this dissertation, we mainly research the reconstruction algorithms of synapses, mitochondria and neurons in the microscopic connectome. In view of the shortcomings of the existing methods, we propose deep learning networks appropriate for the morphology characteristics of the aforementioned targets, which break through the technical bottlenecks in automatic identification of ultrastructure and dense segmentation of neurons, it also reduces the dependence on labor while improving the accuracy and efficiency. The main contributions are listed as follows:

 

1. To tackle the low accuracy and low efficiency of synapses reconstruction algorithm under the complex background, an effective coarse-to-fine automated pipeline is proposed for synapses reconstruction. The proposed method first adopts the detection network based on deep learning to detect synapses and uses the z-continuity of synapses to reduce false positives. Subsequently, it combines the Dijkstra algorithm with the GrabCut algorithm to obtain the segmentation of synaptic clefts. The experimental results in ATUM-SEM and FIB-SEM datasets demonstrate the effectiveness and efficiency of the proposed method, and the average precision of our detection (92.8% in anisotropy, 93.5% in isotropy) and segmentation (88.6% in anisotropy, 93.0% in isotropy) suggests that our method achieves state-of-the-art results. In addition, we utilize this method to obtain the 3D reconstruction and morphological statistics of synapses in large volume EM data, which provide reliable analytical data for neurologists to research synapses.

 

2. To tackle the variety of target structures and noise in mitochondria reconstruction, the method based on 3D supervised full convolution network is proposed to reconstruction mitochondria. In this method, the 3D convolution network could effectively utilize the three-dimensional information of the EM data. The structure combines U-Net and residual convolution module is used to fully integrate the features of different scales so as to identify mitochondria of different shapes and sizes. The strategy of deep supervision is helpful to avoid gradient vanishing in training process. The experimental results on anisotropic and isotropic EM volumes demonstrate the effectiveness of the proposed method. On this basis, we obtain the 3D reconstruction and morphological statistics of mitochondria in large EM data, and we reconstruct the mitochondria in the cell body and dendrites of mouse motor cortex. The reconstruction results prove that mitochondria are connected by ``mitochondrial conduits" in dendrites and provide neurologists a new way to study the function of mitochondria in neurons.

 

3. For the problems of complex background and distinct difference of neuronal structure in neuron reconstruction, we propose an effective neuron reconstruction method based on context residual network. It utilizes multi-scale contextual cues to distinguish neuron membranes from other organelles. The sub-pixel convolution layer in the network is used to reduce the loss of neuron membrane information in up-sampling process. Furthermore, Lifted Multicut is utilized to obtain the 3D segmentation results of neurons. The proposed method now ranks 2/171 in the ISBI EM segmentation challenge and achieves the best result in the mouse piriform cortex dataset. We then apply this method to dense reconstruct a volume of 33*30*40 um3 EM data from mouse cortex and obtain the most of the neuronal structure with a small amount of manual verification.

Pages133
Funding ProjectNational Natural Science Foundation of China[11771130] ; National Natural Science Foundation of China[61673381] ; National Natural Science Foundation of China[61871177] ; National Natural Science Foundation of China[31472001] ; Special Program of Beijing Municipal Science & Technology Commission[Z161100000216146] ; Special Program of Beijing Municipal Science and Technology Commission[Z181100000118002] ; Scientific Instrument Developing Project of Chinese Academy of Sciences[YZ201671] ; Strategic Priority Research Program of the CAS[XDB02060001]
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/23858
Collection类脑智能研究中心_神经计算及脑机交互
Recommended Citation
GB/T 7714
肖驰. 基于深度学习的微观脑连接图谱重建方法研究[D]. 北京. 中国科学院大学,2019.
Files in This Item:
File Name/Size DocType Version Access License
基于深度学习的微观脑连接图谱重建方法研究(15060KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[肖驰]'s Articles
Baidu academic
Similar articles in Baidu academic
[肖驰]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[肖驰]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.