Figure 2.

Reconstruction errors were simulated by applying one of four perturbation models to a synthetically generated ground truth (GT) network. Perturbation models (see Table 3) introduced errors by (A) deleting a percentage of synapses from the overall network, (B) probabilistically inserting synapses where two neuron membranes are closely apposed, (C) probabilistically splitting neurons where process diameter is small, or (D) probabilistically merging neurons where two neuron membranes are closely apposed. These plots show how local NRI scores of individual neurons vary as a result of the introduction of these errors. Several perturbation metrics were used to compare perturbation magnitude to NRI scores. For synapse deletions, neuron NRI scores are compared to the fraction of synapses that were deleted from a given GT neuron. For synapse insertions and neuron merges, NRI is compared to the fraction of synapses not found on the GT neuron. In the case of neuron merges, this means that if neuron A is merged with neurons B and C in the reconstruction, then the perturbation score for neuron A is $\frac{n_B+n_C}{n_A+n_B+n_C}$ where n_A, n_B, and n_C are the number of synapses associated with neurons A, B, and C, respectively. For neuron splits, neuron NRI is compared to the entropy of the synapse distribution across the split pieces of the GT neuron (normalized by the total number of synapses). The color of each neuron's data point indicates the global network in which the neuron resided, and the NRI score for that global network is indicated in the plot's legend. For example, for synapse deletions in plot A, the data points colored dark blue at the bottom right of the plot are neurons from a single perturbed network whose network NRI score is 0.077. Individual neuron NRI scores are close to the network NRI score in this particular case. To enhance visibility, results from only 10% of the neurons (uniformly chosen across neuron size) are plotted. Overlap of markers of different colors is indicative of a broad range of neuron NRI scores for a single reconstructed network.