Automated Synaptic Connectivity Inference For Volume Electron Microscopy
Introduction
The human brain is an intelligent, sophisticated machine. This analogy is accurate in some aspects, and offers an approach to researches related to our brain. Our brain can be divided into four parts: frontal, parietal, temporal and occipital lobes. One criteria for the division is its functionality, or what functions the area is in charge of. For example, the temporal lobe is usually associated with auditory processing and olfactory, while the occipital lobe is usually related to visual information processing.
However, most of the neural behaviors in the brain are very complicated, involving several brain areas to different extents. And different functionalities are not restricted to specific brain areas either. Ambiguity exists everywhere. Thus, when a brain-related disease develops, and functionality deficiencies appear, it is rather difficult to investigate the underlying reason from a macroscopic perspective. Going back to the machine analogy, scientists are now wondering if they could solve the “ambiguity” by approaching on a microscopic-level: the connections among basic units of the brain — neurons. A connectome is a comprehensive map of neural connections in the brain, displaying how neurons wire together and contribute to different functions.
Volume Electron Microscopy
Volume electron microscopy (volume EM) is a commonly used technique for neural circuit reconstruction. In volume EM, a 3-dimensional EM imaging of volumes of the brain can be used to reconstruct details of neuronal shape and connectivity. Differentiations of volume EM were initially developed for examination of the central nervous system (CNS). As mentioned in the introduction, many neurodegenerative diseases could not be traced in an top-to-down fashion. Thus it is necessary to analyze axons, dendrites, and individual synaptic activities with sufficient resolution.
Compared to fluorescence-based labeling approaches that are usually adapted in tissue examination, standard EM stains are not limited by requirements of sparse labeling or super-resolution optical imaging. These stains could result in a relatively unbiased staining of all membrane and synapses. Thus, volume EM could be used to model the complete presynaptic and postsynaptic connectivities of a neuron. It is also a standard operation for all the neurons in the volume, which allows us to build a comprehensive wiring diagram, or a connectome of the brain.
Quantitative methods have become increasingly important in recent years as technology in data processing advances. Volume EM could offer previously unobtainable insights into neuronal computation by offering anatomical circuit reconstruction from large data sets. Technical advancements in volume EM, and increasing computational power have already enabled data sets of sufficient size to reconstruct complete neuronal microcircuits. These new findings have given support to several studies and demonstration how anatomical circuit reconstruction can provide previously unobtainable insights into neuronal computation.
Synaptic Connectivity Inference Pipeline (SyConn)
Nervous systems in vertebrate and invertebrates are densely packed with interweaving neurons, with their axons, dendrites and synapses connecting or overlapping to each other. Thus. trying to hack the detailed connections among thousands of neurons is not an easy job. The connectome, reconstructed from large data sets acquired from volume EM, is a high-dimensional network, which means its analysis requires huge amount of time and effort. Although technology advancements have helped us solved the problem of acquiring sufficient data with good resolution, its analysis still remain a problem. As shown in Figure.1, manual analysis takes millions of hours if we want to reconstruct full details.
It is thus necessary to develop a method that could automatically analyze all the available data to make the construction of a connectome more feasible. In this paper, researchers introduce an automated synaptic connectivity inference pipeline (SyConn) that requires generated neurite skeletons and classifier training data as inputs, and delivers a richly annotated wiring diagram, or components of a connectome. In this inference pipeline, skeletons are converted into volume reconstruction at the first step, followed by synapses and other ultrastructural objects in image data, such as vesicles and mitochondria. Detection of ultrastructural can further augmenting neurites reconstructions.
SyConn framework uses deep convolutional neural networks and random forest classifiers to produce a richly annotated synaptic connectivity matrix through automatic identification of mitochondria, synapses and their cell types. A high-level convolutional neural network (CNN) library that efficiently uses graphics processing units (GPUs) for computing, ElektroNN, was developed specifically to integrate into SyConn. By eliminating the redundant computations and sparse training labels, ElektroNN is optimized for short model training times and fast inference on large data sets.
To convert skeletons into volume reconstructions, a recursive 3D CNN model is trained to detect barrier regions (membrane and extracellular space, ECS) between neurites. ECS could then be utilized to prepare samples for segmentation. Instead of using the area of contact between two neurons as the criterion of whether they connect to each other, researchers chose to detect vesicle clouds and mitochondria together with synaptic junctions. These ultrastructural objects are presented in abundance in pre- and post-synaptic neurons since they act as important factors of information transportation among neurons. Thus detection of co-occurence of vesicle clouds and mitochondria is a good indication of connectivity. Technically, a multi-class CNN was trained to process this step.
It is important to note that there exists a dependence of reported best scores on test set sizes. The multi-class CNN produces great results for small test volumes, possibly because the number of connectivities of these volumes can still be handled. Although the performance displayed in the experiments is promising, it is not sure whether the performance could be extrapolated towards larger data sets, for they could have great diversity.
Based on the previously detected ultrastructural objects, SyConn could further refine the reconstruction by assigning the comparative location of these objects to the neurite hull. This process could assist the classification of subcellular parts and neuronal cell types. In the article, researchers incorporated a random forest classifier (RFC) to classify parts of dendrites as belonging to spine heads, necks or the dendritic shaft. Augmented cell reconstruction is necessary in cell-type identification, which plays a role in structuring connectivity matrix and the following analysis. By comparing mitochondria and vesicle cloud volume density along the neurites, researchers found that those types of neuron with the highest firing rates had the highest densities. Study of ultrastructural objects of neurons and related firing rates might provide insights into their physiological properties in vivo before chemical fixation.
Discussion
Connectomics has gone through fast developments in recent years. Dense connectomic analysis is restricted by the annotation time for synapses and following steps of circuit analysis. SyConn is a great method to cut off analyzing time significantly with a low error rates such that manual proofreading is not necessary (within acceptable errors). For the cases that the data set quality constrains SyConn’s performance, manual check would benefit the accuracy. From the results, we can also see that deep CNNs only require minimal training data to extract ultrastructural information by using pre-trained networks and post-training.
Although the automation improves efficiency notably, automated neurite reconstruction, which is of greater variability and complexity, has not been touched. As we are still in the stage that experts play an important role in biological data analysis, we could foresee that they would have less and less influence in the future. Instead, machines that learned all the rules will probably take the work. Do you think experts in the field would be substituted by computers completely?
Reference:
https://www.csuchico.edu/~pmccaffrey/syllabi/CMSD%20320/362unit4.html
http://www.sciencedirect.com/science/article/pii/S0968432814000250
http://www.sciencedirect.com/science/article/pii/S0959438811001887
Author: Yuka Liu | Editor: Junpei Zhong| Localized by Synced Global Team: Xiang Chen
