Search

The neurosciences have developed methods that outpace most other biomedical fields in terms of acquired bytes. We review how the information content and analysis challenge of such data indicates that electron microscopy (EM)-based connectomics is an especially hard problem. Here, as in many other current machine learning applications, the need for excessive amounts of labelled data while utilizing only a small fraction of available raw image data for algorithm training illustrates the still fundamental gap between artificial and biological intelligence. Substantial improvements of label and energy efficiency in machine learning may be required to address the formidable challenge of acquiring the nanoscale connectome of a human brain.

INTRODUCTION The brain of mammals consists of an enormously dense network of neuronal wires: the axons and dendrites of nerve cells. Their packing density is so high that light-based imaging methods have so far only been able to resolve a very small fraction of nerve cells and their interaction sites, the synapses, in mammalian cortex. Recent advances in three-dimensional (3D) electron microscopy allow researchers to image every nerve cell and all chemical synapses in a given piece of brain tissue, opening up the possibility of mapping neuronal networks densely, not just sparsely. Although there have been substantial advances in imaging speed, the analysis of such 3D image data is still the limiting step. Therefore, dense reconstructions of cortical tissue have thus far been limited to femtoliter-scale volumes, keeping the systematic analysis of axons, neuronal cell bodies and their dendrites of different types, and the dense connectome between them out of reach. RATIONALE Image analysis has made decisive progress using artificial intelligence–based methods, but the resulting reconstructions of dense nerve tissue are still too error-prone to be scientifically meaningful as is. To address this, human data analysis has been integrated into the generation of connectomes and it is the efficiency of this human–machine data analysis that now determines progress in connectomics. We therefore focused on efficiency gains by: (i) improving the automated segmentation quality, (ii) analyzing the automated segmentation for locations of likely errors and directing the human work to these locations only, and (iii) optimizing human data interaction by helping annotators to immediately understand the problem to be solved, allowing fast, in-browser parallel data flight, and by minimizing latency between annotator queries. With this, close to 100 student annotators solved hundreds of thousands of reconstruction problems within just 29 s each, including all preparation and transition time. RESULTS We reconstructed 2.7 m of neuronal wires densely in layer 4 of mouse somatosensory cortex within only ~4000 invested human work hours, yielding a reconstruction ~300 times larger than previous dense cortical reconstructions at ~20-fold increased efficiency, a leap for the dense reconstruction of connectomes. The resulting connectome between 6979 presynaptic and 3719 postsynaptic neurites with at least 10 synapses each, comprising 153,171 synapses total, was then analyzed for the dense circuit structure in the cerebral cortex. We found that connectomic data alone allowed the definition of inhibitory axon types that showed established principles of synaptic specificity for subcellular postsynaptic compartments, but that at scales beyond ~5 μm, geometric predictability of the circuit structure was low and coarser models of random wiring needed to be rejected for dense cortical neuropil. A gradient of thalamocortical synapse density along the cortical axis yielded an enhanced variability of synaptic input composition at the level of single L4 cell dendrites. Finally, we quantified connectomic imprints consistent with Hebbian synaptic weight adaptation, obtaining upper bounds for the fraction of the circuit that could have undergone long-term potentiation. CONCLUSION By leveraging human–machine interaction for connectomic analysis of neuronal tissue, we acquired the largest connectome from the cerebral cortex to date. Using these data for connectomic cell-type definition and the mapping of upper bounds for the learned circuit fraction, we establish an approach for connectomic phenotyping of local dense neuronal circuitry in the mammalian cortex, opening the possibility for the connectomic screening of nervous tissue from various cortices, layers, species, developmental stages, sensory experience, and disease conditions.

The dense circuit structure of the mammalian cerebral cortex is still unknown. With developments in 3-dimensional (3D) electron microscopy, the imaging of sizeable volumes of neuropil has become possible, but dense reconstruction of connectomes from such image data is the limiting step. Here, we report the dense reconstruction of a volume of about 500,000 μm3from layer 4 of mouse barrel cortex, about 300 times larger than previous dense reconstructions from the mammalian cerebral cortex. Using a novel reconstruction technique, FocusEM, we were able to reconstruct a total of 0.9 meters of dendrites and about 1.8 meters of axons investing only about 4,000 human work hours, about 10-25 times more efficient than previous dense circuit reconstructions. We find that connectomic data alone allows the definition of inhibitory axon types that show established principles of synaptic specificity for subcellular postsynaptic compartments. We find that also a fraction of excitatory axons exhibit such subcellular target specificity. Only about 35% of inhibitory and 55% of excitatory synaptic subcellular innervation can be predicted from the geometrical availability of membrane surface, revoking coarser models of random wiring for synaptic connections in cortical layer 4. We furthermore find evidence for enhanced variability of synaptic input composition between neurons at the level of primary dendrites in cortical layer 4. Finally, we obtain evidence for Hebbian synaptic weight adaptation in at least 24% of connections; at least 35% of connections show no sign of such previous plasticity. Together, these results establish an approach to connectomic phenotyping of local dense neuronal circuitry in the mammalian cortex.

Nerve tissue contains a high density of chemical synapses, about 1 per µm3 in the mammalian cerebral cortex. Thus, even for small blocks of nerve tissue, dense connectomic mapping requires the identification of millions to billions of synapses. While the focus of connectomic data analysis has been on neurite reconstruction, synapse detection becomes limiting when datasets grow in size and dense mapping is required. Here, we report SynEM, a method for automated detection of synapses from conventionally en-bloc stained 3D electron microscopy image stacks. The approach is based on a segmentation of the image data and focuses on classifying borders between neuronal processes as synaptic or non-synaptic. SynEM yields 97% precision and recall in binary cortical connectomes with no user interaction. It scales to large volumes of cortical neuropil, plausibly even whole-brain datasets. SynEM removes the burden of manual synapse annotation for large densely mapped connectomes. DOI: http://dx.doi.org/10.7554/eLife.26414.001, eLife digest Each nerve cell in the brain of a mammal communicates with about 1,000 other nerve cells in a complex network. Nerve cells ‘talk’ to each other via structures called synapses that connect the nerve cells together. The number of synapses in the brain is enormous – for example, a human brain contains about one quadrillion synapses. One technique that can be used to look at the synapses in the brain is called 3D electron microscopy. The huge number of synapses in an image makes it impractical for researchers to manually label them. However, current methods that use computers to automatically label synapses work most accurately only on images that are so detailed that they cover only very small volumes of the brain (much less than 1 cubic millimeter). Staffler et al. have now developed a new method, called SynEM, that makes it possible for computers to do all the work of finding the synapses in larger volumes of the brain. Without any input from researchers, SynEM can correctly identify connections between nerve cells 97% of the time, which is far more successful than any other current computer-based approach. Importantly, SynEM also automatically indicates which nerve cells are connected by a given synapse, providing a map of “who talks to whom” across the brain. Together with SynEM, methods that track the cable-like structures (called neurites) that nerve cells grow to find other nerve cells are already allowing us to map the communication networks in the brain. In the far future Staffler et al. hope that such mappings will become so routine that entire human brains could be studied, perhaps to investigate how diseases affect them. DOI: http://dx.doi.org/10.7554/eLife.26414.002