基于深度学习的脑微观结构重建

为了研究脑运行机制以实现类脑智能,我们引入深度学习工具系统解决突触级脑微观重建中大数据自动分析的难题,包括:密集神经元重建、单根神经元追踪、关键细胞器检测和重建.其中我们使用带有候选区域(region proposal network,RPN)的全卷积网络(fully convolutional networks,FCN)检测线粒体和突触,结合SPPUnet(spatial pyramid pooling U-net)网络框架和MultiCut算法进行神经元的重建,各部分在量化分析和视觉上取得了较好的结果.重建工作为学术界开展高通量的突触尺度脑微观结构重建提供有效支持.

Brain microstructure reconstruction based on deep learning

To research the operation mechanism of brain for realizing brain-inspired intelligence, we employ the deep learning tool system to solve the problems of automatic analysis of big data in synaptic-scale brain microstructure reconstruction. The problems include dense neurons reconstruction, single neuron tracing, key organelles detection and reconstruction. Meanwhile, we adopt the fully convolutional networks (FCN) with a candidate region proposal network (RPN) to detect mitochondria and synapses, and integrate deep SPPUnet (spatial pyramid pooling U-net) framework with multi-cut algorithm for neurons reconstruction. Preferable performance have achieved in visual and quantitative analysis. These works provide effective support for high-throughput synaptic-scale brain microstructure reconstruction for neuroscientists.