The neuron reconstructions were taken from a songbird basal ganglia data set and consisted of flood-filling network (FFN)-created supervoxels 25 (SVs), which were agglomerated to sets of super-SVs (SSVs) 8, each corresponding to a single neuron. 1a, b) that are then analyzed either individually or in combination by a CMN. To address this problem, we used a sampling algorithm that homogeneously probes an entire cell at many locations (“Methods” Fig. 18 rendered views of an entire object, potentially sacrificing crucial detail when applied to reconstructions of entire neurons (or very large objects in general), which can have processes as thin as 50 nm extending over millimeters 24. Third, we identify neuronal cell types and compartments, outperforming methods with hand-crafted features based on skeleton representations on the same data, and finally perform high-resolution cell surface segmentation to identify dendritic spines.ĬMNs are convolutional neural networks (CNNs) optimized for the analysis of multi-channel 2D projections of cell reconstructions, inspired by multi-view CNNs for the classification of objects fitting into projections from one rendering site 18, 19. Second, we apply CMNs to the supervised classification of glia cells and use these data to demonstrate the feasibility and effectiveness of a simple top–down false merger resolution strategy. First, we demonstrate that CMNs can be used to automate morphology feature extraction itself by inferring low-dimensional embeddings, dubbed Neuron2vec, through unsupervised triplet-loss training 22, 23. Here we present cellular morphology neural networks (CMNs), which use multi-view projections to enable the supervised and unsupervised analysis of cell fragments of arbitrary size while retaining high resolution. Interestingly, projection-based models often appear to outperform 3D architectures, possibly because of higher-resolution input due to decreased model complexity 18. Outside of the field of connectomics, many approaches have been developed for automated 3D shape analysis, including multi-view two-dimensional (2D) projection-based neural network models 18, 19 as well as voxel- and point cloud-occupancy-based 3D networks 20, 21. Cell types and other biological properties that can be inferred from the morphology of a neuron are required for the interpretation of a connectome 11 and can also constrain automatic neuron reconstruction itself 17. These advances have enabled automatic morphology analyses on the neuron (fragment) scale, which were previously restricted to direct segmentation error detection 5, 14 or the use of manual skeletons with data-specific hand-crafted features 11, 15, 16. Multi-beam scanning electron microscopes 2 and transmission electron microscopes equipped with fast camera arrays can now generate data sets exceeding 100 TB 3, a development that was fortunately accompanied by substantial progress in neuron reconstruction 4, 5, 6, 7, 8, 9 and the automatic analysis of synapses 10, 11, 12, 13. Finally, we demonstrate that CMNs can be used to identify subcellular compartments and the cell types of neuron reconstructions.Īdvances in volume electron microscopy (VEM) have led to increasingly large three-dimensional (3D) images of brain tissue, making manual analysis infeasible 1. Using unsupervised training, we infer morphology embeddings (Neuron2vec) of neuron reconstructions and train CMNs to identify glia cells in a supervised classification paradigm, which are then used to resolve neuron reconstruction errors. Here, we introduce cellular morphology neural networks (CMNs), based on multi-view projections sampled from automatically reconstructed cellular fragments of arbitrary size and shape. However, methods for automating the morphological analysis of nanometer-resolution reconstructions are less established, despite the diversity of possible applications. Significant progress has recently been made in automated neuron reconstruction as well as automated detection of synapses. Reconstruction and annotation of volume electron microscopy data sets of brain tissue is challenging but can reveal invaluable information about neuronal circuits.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |