Scalable Software for Reverse Engineering Neural Circuits from Histology

描述（由申请人提供）：据估计，人脑具有大约1000亿个神经元，这些神经元通过超过10万英里的轴突和四分之一的突触连接（约10^15或2^50个连接）连接。相比之下，仅波士顿市的人类大脑中的突触连接就比世界上所有甜点和海滩上的沙粒还要多（~10^20）。每个大脑中的神经回路被称为连接体，了解它如何工作并实现认知，意识或智力是科学中最基本的问题之一。 考虑到这种复杂性，即使是最简单的行为背后的神经回路也不为人所知，这并不奇怪。直到最近，试图充分描述这种电路甚至从来没有认真对待，因为它是 被认为在数字上过于复杂。 然而，在过去五年中，脑组织的制备、切片和成像方面的现代进步使生物学家能够以高度自动化的方式在仅几纳米的尺度上对神经连接进行成像。 神经科学研究人员正在使用共聚焦和电子显微镜技术，以高分辨率成像连续切片。当前的三维图像数据集的大小高达几TB。 通过自动化和更快的成像，我们预计 数据集的大小以数量级增加。 不幸的是，处理和分析这些图像以识别任何哺乳动物大脑的连接体仍然是一项令人难以置信的艰巨任务，世界上只有少数几个小组已经开始解决 这个问题 我们建议开发必要的计算基础设施映射神经元的布线在大量的神经组织，已被切割成连续切片。 我们将开发一个开放源代码的系统，支持分析任意大的图像体积.通过能够在合理的时间内（几天或几周而不是几年）跟踪体积中的每个神经过程，我们的系统将能够协同努力开发有效的自动分割和跟踪方法。 拟议的系统将支持远程数据访问，以便地理位置不同的研究小组可以同时访问庞大的数据集。 将开发定制客户端来实现各种分割算法，并将结果上传到中央数据库。 以这种方式，可以将用一种算法获得的分割结果与用相同数据集上的另一种算法获得的分割结果进行比较。我们还将实现融合方法，该方法将不同算法的分割结果作为输入，并通过将分割结果从一个部分链接到下一个部分来生成神经过程的跟踪。

DESCRIPTION (provided by applicant): A human brain is estimated to have roughly 100 billion neurons connected through more than 100 thousand miles of axons and a quadrillion of synaptic connections (~10^15 or 2^50 connections). As a comparison, there are more synaptic connections in human brains in the city of Boston alone than grains of sand in all the desserts and beaches in the world (~10^20). The neural circuit within each brain is called its connectome, and understanding how it works and enables cognition, consciousness, or intelligence is one of the most fundamental questions in science. Given this complexity it is not surprising that the neural circuits underlying even the simplest of behaviors are not understood. Until recently, attempts to fully describe such circuits were never even seriously entertained, as it was considered too numerically complex. However, modern advances in the preparation, sectioning, and imaging of brain tissue in the last five years have enabled biologists to image neural connectivity at scales of only a few nanometers in a highly automated manner. Neuroscience researchers are using confocal and electron microscopy techniques to image serial sections at high resolution. Current three-dimensional image datasets are up to several terabytes in size. With automation and faster imaging, we expect dataset sizes to increase by orders of magnitude. Unfortunately, processing and analyzing these images in order to identify the connectome of any mammalian brain is still an incredibly difficult task, and only a few groups across the world have started to address this problem. We propose to develop the computational infrastructure necessary for mapping the wiring of neurons in a large volume of neural tissue that has been cut into ultrathin serial sections. We will develop an open- source system that supports analysis of arbitrarily large image volumes. By being able to trace every neural process in a volume within a reasonable amount of time (days or weeks instead of years), our system will enable a collaborative effort to develop efficient automatic methods for segmentation and tracing. The proposed system will support remote data access so that the enormous datasets can be accessed simultaneously by geographically diverse research groups. Custom clients will be developed to implement various segmentation algorithms, with results uploaded to a central database. In this way the segmentation results obtained with one algorithm can be compared against those obtained with another algorithm on the same datasets. We will also implement fusion methods that will take as input the segmentation results from different algorithms and that will generate the tracings of neural processes by linking segmentation results from one section to the next.