动物实验 z-bo 2020-09-02 457

　　The human brain is an intelligent and complex machine. This analogy is accurate in several ways and provides a method for the field of brain research. As we all know, the human brain is divided into four parts: frontal lobe, parietal lobe, temporal lobe and occipital lobe. One of the criteria in this classification is the function or type of function the area is responsible for. For example, the temporal lobe is often associated with auditory processing and smell, and the occipital lobe is often associated with visual information processing. However, most neurobehaviors in the brain are very complex and involve multiple areas of the human brain to varying degrees. Its function is not limited to the division of specific brain regions. Ambiguity is everywhere. Therefore, when brain diseases and dysfunction occur, it is difficult to study the root cause from a macro level. Returning to the machine analogy, scientists now want to understand whether we can eliminate this "ambiguity" at the micro level through the connections between the basic units of brain neurons. The connection group is a panoramic view of the neural connections in the brain, showing the connections of neurons and how they drive various functions. Volume electron microscopy (VolumeEM) is a common technique used to reconstruct neural circuits. Among them, the three-dimensional EM imaging technology of brain volume can be used to reconstruct detailed information about the shape of neurons and their connections. The difference in volumeEM development began with the need for central nervous system (CNS) testing. As mentioned at the beginning, many neurodegenerative diseases cannot be tracked through a top-down approach. Therefore, a sufficiently large resolution must be used to analyze axons, dendrites and individual synaptic activity. Generally, fluorescent labels are suitable for histology, but standard EM dyes are not limited to sparse labels or super-resolution optical imaging requirements. These pigments may stain all cell membranes and synapses fairly evenly. Therefore, VolumeEM can be used to form a complete connection between the pre-synaptic structure and the post-synaptic structure of a specific neuron. This standard operation can be extended to all neurons in the brain volume, so you can build a comprehensive brain nerve wiring diagram or a set of brain connections. In recent years, with the development of data processing technology, quantitative methods have become more and more important. By reconstructing anatomical circuits from large data sets, VolumeEM can provide insights not previously available in neural computing. The technological advancement and computing power of VolumeEM make it possible to reconstruct a complete neural microcircuit with a sufficiently large data set. These new discoveries have provided support and demonstrations for multiple research projects. Anatomical circuit reconstruction technology can provide insights previously unobtainable in neural computing. Synaptic Connection Reasoning Conduit (SYCONN) The vertebrate and invertebrate nervous system is tightly entwined with neurons, and their axons, dendrites and synapses are connected or overlapped with each other. I'm. Therefore, decrypting the connection details between thousands of neurons is not an easy task. Rebuilt from the large data set of VolumeEM, this connection group is a high-dimensional network. In short, this analysis requires a lot of time and effort. Advances in technology have provided sufficient data and sufficient resolution, but the analysis process is still a problem. As shown in Figure 1, manual analysis requires millions of hours to reconstruct all the details.

　　Therefore, in order to make the construction of connection groups more feasible, we need to develop a method that can automatically analyze all available data. In this article, the researchers introduced an automatic synaptic connection inference process (SyConn). The model must use the generated neurite bones and classifier training data as input, and provide a large amount of annotation information for the connection diagram or connection group components for the worksheet. In this reasoning process model, the first step is to reconstruct the neurite skeleton. Then, it transforms the synapses and other hyperfine objects in the image data, such as vesicles and mitochondria. Ultrastructure detection can further enhance the effect of neurite remodeling.

　　SyConn framework uses deep convolutional neural network and Random Forest Classifier (Random Forest Classifier, RFC) to automatically identify mitochondria, synapses and other cell types, as well as a rich synaptic connection matrix with annotation information. produce. ElektroNN is an advanced convolutional neural network (CNN) library that can effectively use graphics processing units (GPU) for calculations, especially when integrated into SyConn. By eliminating redundant calculations and sparse training tags, ElektroNN optimizes model training time and inference speed for large data sets. To

　　Skeleton is converted to volume reconstruction, we can train a recursive 3D convolutional neural network model to detect the barrier area between neurites (membrane and extracellular space, ECS). You can then use ECS to prepare samples for segmentation. The researchers chose to test the vesicle cloud and mitochondria together to measure whether two neurons are connected to each other, rather than two neurons touching each other. These superstructure objects are abundant in pre-synaptic and post-synaptic neurons because they are important signal components between neurons. Therefore, detecting the co-existence of vesicle clouds and mitochondria is a related good sign. Technically speaking, multiple types of convolutional neural networks have been trained to handle this step.

　　Please note that the best training results you get from these reports depend on the size of the test set. Multi-class convolutional neural networks can work well on small test sets. This may be because the number of connections for the volume is still manageable. The performance shown in the experiment has a bright future, but I am not sure whether it can be generalized to larger data sets, but in the end, the large data sets vary greatly. Based on the previously detected hyperfine structure objects, SyConn can further improve the reconstruction model by assigning its relative positions to different parts of the neurites. This process helps to classify the intracellular parts and nerve cell types. In this article, the researchers added a random forest classifier to classify dendritic spines, dendritic spines or dendritic spines. Unit type recognition requires enhanced unit reshaping, and this recognition procedure plays an important role in the construction of the connection matrix and subsequent analysis. By comparing the volume density of mitochondria and vesicle clouds on neurites, the researchers found that the neuron type with the highest firing frequency had the highest density. The study of neuron ultrastructure and related release rate may provide some insights into the physiological characteristics of life before chemical fixation.

　　Discussion

　　Connectomics (Connectomics) has made great progress in recent years. The dense connection hybrid analysis is limited by the synaptic annotation time and follows the circuit analysis procedure. SyConn can reduce the error rate (within the acceptable error range) and significantly reduce analysis time, eliminating the need for work calibration. For the problem that the quality of the data set limits the performance of SyConn, manual inspection can improve the accuracy. The results also show that, using well-trained networks and post-training procedures, superconvolutional neural networks require minimal training data to extract information about hyper-fine structured objects.