Fig. 4: Whole-brain 3D registration workflow and results.
From: D-LMBmap: a fully automated deep-learning pipeline for whole-brain profiling of neural circuitry
a, The multiscale learning-based framework with multi-constraints for the whole-brain 3D registration. The registration input contains multiple sources, including the LSFM experimental brain, the style-transferred experimental brain, the segmented brain regions and the reference atlas and atlas brain regions. All the inputs are downsampled twice and then computed with rigid and affine transformation using convolutional layers. A neural network based on VoxelMorph is used for deformable transformation. b, Registration results of representative brains from different modalities, including the LSFM brain in the autofluorescence channel, the MRI brain and the LSFM brain in the stained-specific channel. Orange lines indicate where the brain region boundaries are defined in the Allen atlas overlaid. (Scale bar, x, y, z = 1 mm). c, Quantitative evaluation and methods comparison of whole-brain registration on 12 LSFM brains in the autofluorescence channel. Median Dice score of six individual major brain regions (CP, HPF, CTX, CB, CBX and BS) (left). Average median Dice score across six brain regions (Right). Box plot: center line, median; box limits, upper and lower quartiles; whiskers, 1.5× interquartile range; points, individual data points. d, Quantitative evaluation and methods comparison of whole-brain registration on small brain structures using nine LSFM brains. Median Dice score of five small brain structures (act, fr, mtt, IPN, Hb) (left). Average median Dice score across five small brain structures (right). Box plot, center line, median; box limits, upper and lower quartiles; whiskers, 1.5 × interquartile range; points, individual data points. e, Evaluating the whole-brain registration by the false-positive report of cell distribution in the ventricular system. (i) Representative horizontal and coronal brain slices containing the LV, 3rd V and AQ; (ii) Numbers of cells that are falsely reported to be in LV, 3rd V and AQ by ClearMap, mBrainAligner and D-LMBmap (one-way analysis of variance followed by Dunnett’s multiple comparison test. F (1.003, 2.006) = 30.44, P = 0.031, n = 3). *P < 0.05. Measure of center, mean; error bars, mean ± s.d.
