Science cover paper, 3981 hours to rebuild 500,000 cubic micron mouse brains, artificial neural network landmark study!

Recently, researchers from the Max Planck Institute of brain in Germany have reconstructed the morphological characteristics of 89 neurons and their connections in the barrel cortex of mice with high spatial resolution by using advanced automated imaging and analysis tools.the volume of about 500000 cubic microns was reconstructed from the fourth layer of the barrel cortex of mice, which was about 300 times larger than the previous intensive reconstruction from mammalian cerebral cortex.this achievement is on the cover of this issue of science.the mammalian brain, in terms of the number of nerve cells and the density of connections between them, is the most complex network known.mapping neuronal circuits intensively by imaging every synapse and all neurofilaments in brain tissue has always been a major challenge.recently, researchers from Max Planck Institute of brain in Germany have imaged and analyzed the cerebral cortex of mice.by using advanced automated imaging and analysis tools, the morphological characteristics of 89 neurons and their connections in the barrel cortex of mice were reconstructed with high spatial resolution.this achievement is on the cover of this issue of science.the researchers applied optimized AI based image processing and effective human-computer interaction to analyze about 400000 synapses and about 2.7 meters of neuronal cables.they created a connectome between about 7000 axons and about 3700 postsynaptic neurites, which was 26 times larger than that obtained from mouse retinas 60 years ago.it is important that this reconstruction is at the same time larger and more efficient than that applied to the retina.this method revealed information about the connectivity of inhibitory and excitatory synapses in the cortical and excitatory thalamic cortex connections.the researchers reconstructed a volume of about 500000 cubic microns from the fourth layer of the barrel cortex in mice, about 300 times larger than previous intensive reconstructions from mammalian cerebral cortex.connectome data can extract inhibitory and excitatory neuronal subtypes that cannot be predicted by geometric information.the researchers quantified the imprinting of the junction group, which produced an upper bound on the part of the circuit consistent with the long-term enhancement of saturation.these data establish a connectome phenotypic analysis method for the local dense neuronal circuits in mammalian cortex.Alessandro Motta, lead author of the study, said: "some synaptic plasticity models make specific predictions of synaptic weight gain in learning, such as recognition trees or cats, and we are surprised to find such information and accuracy even in a relatively small cortex."next, we will interpret this study for you: the efficiency of human-computer data analysis determines the progress of connectomics. The study proposes three points to improve the analysis efficiency. Through the use of artificial intelligence based methods, image analysis has made key progress, but the reconstruction of dense nerve tissue is still prone to errors, so that it does not show its scientific significance.in order to solve this problem, human data analysis has been integrated into the generation of connectome. Now, the efficiency of this human-computer data analysis determines the progress of connectomics.therefore, researchers focus on how to improve the analysis efficiency, which mainly includes the following points: improve the quality of automatic segmentation; analyze the possible wrong positions in automatic segmentation, and guide the manual work to these positions only; optimize the personnel data interaction by helping annotators, so as to immediately understand the problems to be solved, so as to realize the parallel data in the browser And minimize latency between annotator queries.Advanced microscopy + artificial intelligence: 3981 hours to completely reconstruct each cell of a small mouse brain, researchers used continuous scanning electron microscopy (SBEM) to obtain a 3D EM dataset from the fourth layer of the primary somatosensory cortex of a 28 day old mouse.for dense reconstruction (Fig. 1, e to h), the researchers aligned the images in 3D and applied a series of automatic analysis techniques [Segem, synem, connectem and typeem], and then focused on manual annotation (focusem).the researchers reconstructed 89 neurons in the data set (Figures 1, e and F), which accounted for only 2.6% of the total connection length.Fig. 1: reconstitution of dense neuronal junctions in the fourth layer of primary somatosensory cortex in mice.in order to reconstruct the axons that make up most of the wiring in dense circuits, researchers have adopted a scalable distributed annotation strategy, which can identify the uncertain positions in automatic reconstruction, and then solve the problem through targeted manual annotation.in order to reduce the manual annotation time required, obtain automated refactoring with low error rates, use efficient algorithms to identify locations for centralized manual checks (queries), and minimize the time spent on each query.to this end, as shown in Figure 2a, researchers developed an artificial intelligence based algorithm to evaluate EM image data and convolutional neural network (CNN) filtered image data (Figure 2b).Figure 2: the method of efficient dense junction group reconstruction.using this flexible annotation structure, the reconstruction of 2.69 m dense neuronal processes was achieved (Fig. 1, G and H), and the time of manual work was only 3981 hours.this is 10 times faster than the work of K. Eichler et al. In 2017 to reconstruct the brains of Drosophila larvae, and 20 times faster than the work of scientists in 2013 to completely reconstruct small mammalian retinas.the researchers obtained a set of connections between 34221 presynaptic axons and 11400 postsynaptic processes (Fig. 3).Figure 3: postsynaptic targets and dense cortical junctions.Figure 4: can these local connection rules be derived only from the geometry of axons and dendrites? The researchers first quantified the overall relationship between the spatial distribution of axons and dendrites and the synaptic establishment between them (Fig. 5). Br / > the effect on the geometry of the neural process and the cortical wiring of the < 5 . discussion: Based on the obtained dense brain circuit reconstruction, the research analysis highlights the plasticity model. Using focusem, researchers obtained the first dense circuit reconstruction from mammalian cerebral cortex, which is about 300 times larger than the previous dense reconstruction in cortex. This scale is enough to analyze the axon pattern of subcellular internal axons and give priority to some postsynaptic subcellular compartments Inhibitory axon types can be defined only based on topological information of connections (FIGS. 3 and 4). in addition to inhibitory axons, some excitatory axons also showed this subcellular endocrine preference (Fig. 4). the geometric arrangement of axons and dendrites only explained part of the synaptic internal changes, thus eliminating the rough random model of cortical wiring (Fig. 5). a large number of TC synaptic gradients in L4 enhanced the heterogeneity of synaptic input components at the level of a single cortical dendritic cell (Fig. 6), accompanied by an endogenous decrease in preferential inhibitory inputs from AD. the consistency of synaptic size between axon and dendrite pairs indicates that each part of the circuit is consistent with saturated synaptic plasticity, which sets an upper limit for the "learning" part of the circuit (Fig. 7). focusem enables dense mapping of the circuits in the cerebral cortex with the flux capable of connector screening. in Fig. 7, the composition of synaptic inputs along the L4 dendrite, which is consistent with plasticity, was studied. It was found that the enhanced TC input of L4 excitatory cells and the decreased direct inhibitory input from ad preferentially in (Fig. 6, h to k) can be explained in the inhibition circuit described previously. considering that growth hormone inhibin (SST) positive in preferentially targets ads and preferred corpuscular albumin (PV) positive in, this may mean that SST in based de inhibition can enhance TC input by inhibiting the PV input of peroxides collected by feedforward inhibition, and simultaneously reduce the direct inhibitory component from SST in. in any case, this finding points to a circuit configuration in which TC input variability is enhanced between neurons of the same excitation type in cortical layer 4, and also provides evidence for increased heterogeneity of input composition based on dendritic synapses. plasticity pathway researchers in the connectome group interpreted the synaptic data based on the upper limit of synaptic pairs that may have experienced some plastic models (Fig. 7). although this analysis detected pairs of synapses exposed to saturated plasticity (i.e., a possible plasticity event resulting in the final weight state of two synapses), another explanation is the dynamic circuit, in which only some synapses have shown saturated plasticity at any given point in time, while others (or all) are undergoing plastic changes. the researchers expect that a more refined plasticity model of the whole circuit will also make testable predictions, which can be accessed through connection group snapshot experiments. the milestone research of artificial neural network is expected to promote the global related research plan. The methods and conclusions proposed in this study will open the way for screening the connective tissue of nerve tissue from various cortices, layers, species, developmental stages, sensory experience and disease status. even a small mammalian cortical neuron has a high correlation density, so that the possible "learning" features of the circuit can be extracted. This fact makes this method a promising method to study the structural setting of mammalian nervous system. after nearly ten years of work, researchers are enthusiastic about their achievements. helmstaedter said: "to be able to take a piece of cortex, do the hard work, and then get the whole communication map from that beautiful network is what we've been working on for the past decade. "he also said," the goal of mapping neural networks in the cerebral cortex is a major scientific adventure, and it's also because we want to be able to extract information about how the brain has become such an efficient computer, unlike today's AI. "not only that, this project also hopes to promote the progress of relevant research projects around the world. helmstaedter mentioned the research fields of major participants including Google and the research plan of the American Intelligence Agency (iarpa), "the major global programs hope to learn about the future of artificial neural networks from biological neural networks. we are proud of our first milestone, which is the realization of a dense local cortical connectivity group with the largest public funding from the Planck Institute. ". paper links: - end -