Fig. 2: Deep-learning-based segmentation.

a, Rendering of 85 × 69 × 14-µm³ LICONN volume (native tissue scale) from hippocampal neuropil in CA1, overlaid with dense FFN-based segmentation of neuronal structures in the bottom corner. Neuronal structures were comprehensively proofread in this volume by correction of split and merge errors (without any manual painting of voxels; see https://neuroglancer-demo.appspot.com/#!gs://liconn-public/ng_states/expid82.json for original data and proofread segmentation). Scale bar, 10 μm. b, Magnified view from a single plane with (top to bottom) raw structural data (CLAHE applied), dense segmentation after proofreading and overlay. Scale bars, 1 μm. c, Rendering of 5.8% of axons contained in the volume in a (see https://neuroglancer-demo.appspot.com/#!gs://liconn-public/ng_states/expid82_fig2_axons.json for browsable data). Scale bar, 10 μm. d, Rendering of 27.3% of dendrites and a small number of axons (see https://neuroglancer-demo.appspot.com/#!gs://liconn-public/ng_states/expid82_fig2_dends.json for browsable data). e, Rendering of example dendrites. Scale bar, 10 μm. f, Spatial arrangement of selected axonal and dendritic structures, highlighting various types of contact. Scale bar (top right), 2 μm. g, Segment size (number of voxels on a logarithmic scale) of neuronal structures for the base FFN segmentation (blue), after automated agglomeration (white) and after full manual proofreading of the automated agglomeration (yellow) for the dataset in a (n = 1 dataset). Vertical bar, lower and upper quartiles; dot, median; vertical lines, 1.5× interquartile range. h, Edge accuracy for the base segmentation (blue), after automated agglomeration (white) and after manual proofreading (yellow). i, Distribution of spine-head volumes in the same segmentation volume, analysed for 59,332 spines. Percentage numbers refer to the intervals indicated by the vertical lines (<0.01, 0.01–0.05, 0.05–0.1, 0.1–0.2, >0.2 µm³).

Source Data