Fig. 2: Automated axon annotation and whole-brain axon segmentation.
From: D-LMBmap: a fully automated deep-learning pipeline for whole-brain profiling of neural circuitry
a, A 3D whole brain and the selected 3D cubes containing sparse axons, dense axons and artifacts, respectively. b, The automated annotation workflow for 3D cubes with ‘pure’ axons. It contains two steps, adaptive binarization (second row) and axon skeletonization (third row). c, The introduced data augmentation strategy to improve the diversity of annotated cubes, including CutMix, histogram matching and local contrast augmentation. d, The workflow of the DNN for axon segmentation, including preprocessing for training cube packaging, training strategy self-regulation and network training. e, The comparison of axon segmentation results of example cubes containing sparse axons, dense axons and mixed artifacts and axons between TrailMap and D-LMBmap. The blue squares indicate the regions for zoom-in comparison. (Scale bar, x, y, z = 60 μm). f, Quantitative comparison between D-LMBmap and TrailMap under the evaluation of Dice, ClDice, Precision and ClPrecision (two-tailed paired t-test, n = 10. ClDice, P = 0.0003, t = 5.622, d.f. = 9; ClPrecision, P = 0.0005, t = 5.223, d.f. = 9; ClRecall, P = 0.0622, t = 2.129, d.f. = 9; Dice, P = 0.000095, t = 6.641, d.f. = 9). Measure of center, mean; error bars, mean ± s.d.
