Fig. 5: DeepIsoBrain for isotropic whole-brain imaging and accurately reconstructing long-range projection neurons.
a Online self-learning during whole-brain HR imaging. A data block with 5003 voxels containing dense neurites is sufficient for training Self-Net. b On-demand isotropic restoration during brain-wide single-neuron reconstruction. c The number of raw data cubes (2003 voxel size) of the whole-brain dataset, and the traversed and Self-Net-restored raw data cubes while reconstructing all 20 neurons. d Additional time cost comparison using different isotropic strategies from the current highest of 0.2 × 0.2 × 1 μm3 to 0.2 × 0.2 × 0.2 μm3 for whole-brain imaging. e Comparing the tracing difficulties in 50 randomly-selected challenging cases using the raw and Self-Net data. f The tracing error rate in the challenging regions and the length and branch number of erroneously-traced fibers using the raw and Self-Net data. The tracing GTs for these challenging areas were derived from the reconstruction consensus among three neuroanatomical experts, who independently performed tracing through slice-by-slice editing in the HR lateral views of the raw data. g A typical serious reconstruction error of tracing a neuron in the raw data. The accurately and erroneously-traced axon fibers are colored green and pink, respectively. Scale bar: 1 mm. h Three orthogonal perspectives of the image block indicated by the yellow box in (g) using the raw data and Self-Net-restored data. The dashed lines indicate the tracing results of the same annotator and the colors indicate different branches. The insets show enlarged views of the regions indicated by the magenta arrowheads. The tracing results using Self-Net data were consistent with the gold-standard tracing results. Scale bar: 10 μm
