21The dense circuit structure of the mammalian cerebral cortex is still unknown. With 22 developments in 3-dimensional (3D) electron microscopy, the imaging of sizeable 23 volumes of neuropil has become possible, but dense reconstruction of connectomes 24 from such image data is the limiting step. Here, we report the dense reconstruction 25 of a volume of about 500,000 µm 3 from layer 4 of mouse barrel cortex, about 300 26 times larger than previous dense reconstructions from the mammalian cerebral 27 cortex. Using a novel reconstruction technique, FocusEM, we were able to 28 reconstruct a total of 0.9 meters of dendrites and about 1.8 meters of axons investing 29 only about 4,000 human work hours, about 10-25 times more efficient than previous 30 dense circuit reconstructions. We find that connectomic data alone allows the 31 definition of inhibitory axon types that show established principles of synaptic 32 specificity for subcellular postsynaptic compartments. We find that also a fraction of 33 excitatory axons exhibit such subcellular target specificity. Only about 34 35% of inhibitory and 55% of excitatory synaptic subcellular innervation can be 35 predicted from the geometrical availability of membrane surface, revoking coarser 36 models of random wiring for synaptic connections in cortical layer 4. We furthermore 37 find evidence for enhanced variability of synaptic input composition between neurons 38 at the level of primary dendrites in cortical layer 4. Finally, we obtain evidence for 39 Hebbian synaptic weight adaptation in at least 24% of connections; at least 35% of 40 connections show no sign of such previous plasticity. Together, these results 41 establish an approach to connectomic phenotyping of local dense neuronal circuitry 42 in the mammalian cortex. 43 44 109First, we reconstructed those neurons, which had their cell bodies in the tissue 110 volume ( Fig. 1h,i, Supplementary Material 1, n=125 cell bodies, of these 97 111 neuronal, of these 89 reconstructed with dendrites in the dataset) using a set of 112 simple growth rules for automatically connecting neurite pieces based on the 113 segment-to-segment neighborhood graph and the connectivity and neurite type 114 classifiers ( Fig. 1f, see Methods). As a result, we obtained fully automated 115 reconstructions of the neuron's soma and dendritic processes. Notably with a 116 minimal additional manual correction investment of 9.7 hours for 89 cells (54.5 mm 117 6 dendritic and 2.1 mm axonal path length), the dendritic shafts of these neurons could 118 be reconstructed without merge errors, but 37 remaining split errors, at 87.3% 119 dendritic length recall ( Fig. 1h,i, Supplementary Material 1, see Methods). This 120 reconstruction efficiency compares favorably to recent reports of automated 121 segmentation of neurons in 3D EM data from the bird brain obtained at about 2-fold 122 higher imaging resolution (Januszewski et al., 2018), which reports soma-based 123 neuron reconstruction at an error rate of beyond 100 errors per 66 mm dendritic 124 ...
