Extended Data Figure 4: Automated Segmentation of Neuronal Morphologies using Convolutional Neural Networks (CNNs).

(a) Overview of XNH image volume encompassing the anterior half of the VNC and the first segment of a front leg of an adult Drosophila (200 nm voxels). A smaller, higher resolution (50 nm voxels) volume centered on the prothoracic (T1) neuromere of the VNC and including the initial segment of the leg nerve was used for automatic segmentation. (b) Schematic of U-NET CNN architecture used for automated segmentation (adapted from ⁴¹). Each blue arrow represents two successive convolutions. (c) Morphological comparison of the motor neuron with the largest-diameter branches out of all front leg motor neurons, reconstructed from three different flies using different modalities. Arrows indicate the largest-diameter branches, which match well across the three reconstructions. Left: Reconstruction using automated segmentation of XNH images. Gray segment indicates a merge error that was corrected during proofreading (Methods). Middle: Reconstruction from LM images of a dye-filled motor neuron labeled by 81A07-Gal4. This motor neuron controls the tibia flexor muscle and produces the largest amount of force of any fly leg motor neuron yet identified³⁹. Adapted from Azevedo et al.³⁹. Right: Skeleton reconstruction from EM images. Adapted from Maniates-Selvin et al⁷¹. (d) Population of 90 neurons used for evaluating segmentation error rates. Skeletons were categorized based on their morphologies (as in Fig. 4f) ^40,42,43. White circle indicates the boundary of the T1 neuropil. A, anterior; P, posterior. (e) Examples of merge and split errors. True membrane locations are shown in black. Errors usually result from incorrect prediction of which voxels correspond to membranes. (f) Average error rates of segmentation for the 90 neurons shown in (f). Automated segmentation is parametrized by an agglomeration threshold amounts to a trade-off between split and merge errors. Data points indicate split and merge error rates for different agglomeration thresholds (Methods). The ideal segmentation minimizes the time needed to identify and fix split and merge errors during proofreading (red arrow). Merge error calculations based on comparisons to sparse manual tracing are likely an underestimate of the true number of merge errors. Note that the human-annotated, ground truth segmentation of XNH data excludes some areas where features are too small to resolve; thus these error metrics for XNH segmentation may not be directly comparable to what has been reported for EM. (g-j) Automated segmentation of XNH data in mouse cortex (primary somatosensory, layer 5, 30 nm voxels). (g) Raw data (h) Affinities (zyx corresponding to RGB colors). (i) selected segmentation labels corresponding to (e). (j) Selected 3D renderings of segmented neuron fragments. (k) Large FOV segmentation of myelinated axons in the white matter below mouse parietal cortex. Segmentation of such myelinated axons can enable tracing of long-range inputs between brain areas at single-cell resolution.