Fig. 3: Brain-style transfer and automated brain region segmentation.
From: D-LMBmap: a fully automated deep-learning pipeline for whole-brain profiling of neural circuitry
a, The input consists of LSFM brain slices, while the reference images are from the Allen atlas. The DNN based on CycleGAN consists of two generators and two discriminators. Generator A generates Allen-style images from LSFM-style images, whereas Generator B generates LSFM-style images from Allen-style images. Discriminator A discriminates between original and synthetic LSFM images and Discriminator B does the same for Allen images. The final output is the synthetic Allen images, which are converted to the Allen style with their original content. The brain outlines segmented by CEA-Net are employed as constraints during CycleGAN training. b, Example results showing the input LSFM brain sample has been successfully transferred into a synthetic ‘Allen-style’ brain. c, The network architecture for semi-supervised multiview brain region segmentation. The network is trained by one atlas/annotated brain and several unannotated brains. The 2D brain slices from two views, coronal and horizontal, are extracted for training in Semi-CEA. After training, brain region predictions in coronal and horizontal views are transformed into a unified view and then combined for the computation of consistency loss (MSE loss). The semi-loss and the multiview loss are integrated for the whole deep model training. d, Examples of brain region segmentation results in coronal and horizontal views. Six main brain regions are color coded (CTX, green; CP, blue; HPF, light green; BS, pink; CB, orange; and CBX, yellow). The white lines in the amplified images mark the edge of segmented brain regions. e, Quantitative evaluation of different brain region segmentation methods on autofluorescence channel LSFM brains using the atlas-trained pipeline (n = 12). The brain data and annotations used for training the multiview Semi-CEA deep model (left). Region-wise median Dice score for six brain regions (CP, HPF, CTX, CB, CBX and BS) (middle). Average median Dice score of different methods (right). f, Quantitative evaluation of different brain region segmentation methods on autofluorescence channel LSFM brains using the sample-trained pipeline (n = 8). Box plot, center line, median; box limits, upper and lower quartiles; whiskers, 1.5 × interquartile range; points, individual data points.
