see: Ball et al. 2024
μBrain a 3D digital atlas of the developing brain at micrometre scale based on a public resource of 81 serial histological sections of a prenatal human brain at 21 postconceptional weeks (PCW). Source data included serial coronal sections (20μm thickness) obtained from the right hemisphere of a single prenatal brain specimen, Nissl-stained, imaged at 1 micron resolution and labelled with detailed anatomical annotations.
To correct for tissue artefacts present in histological data, we designed an automated detect-and-repair pipeline for Nissl-stained sections based on pix2pix, a Generative Adversarial Network (GAN):
pix2pix models were implemented in Keras
see: models.py
We trained the pix2pix model on 1000 pairs of 256 × 256 image patches from (downsampled) 20μm resolution Nissl-stained sections and corresponding label annotations.
To perform repair of whole sections, we split each label image into patches of 256 × 256 pixels with an 8 pixel overlap and passed them through the trained generator. The resulting, synthetic Nissl contrast patches were stitched together into a full section matching the dimensions of the original image
To detect regions of the original Nissl-stained section that needed repair, we designed an automated outlier detection method based on the Median Absolute Deviation (MAD) of pixel hue and saturation.
Following automated repair of major tissue artefacts present in the histological data, we aimed to develop a 3-dimensional reconstruction of the fetal brain to facilitate comparison with in vivo MR imaging data.
Using the middle section as a reference, repaired Nissl-stained sections were aligned using a graph-based, slice-to-slice registration. Pairwise rigid transforms were estimated between each section and its neighbouring sections in the direction of the reference. Dijkstra’s shortest-path algorithm was then used to calculate the set of transforms with lowest cost to align a given section to the reference. See also
see: initial_sequential alignment.py
We used a population-based average anatomical image: the 22-week timepoint of the Gholipour et al. spatio-temporal fetal MRI atlas as shape prior for 3D reconstruction. After matching MRI-based tissue labels to the μBrain tissue labels, we upsampled the MRI template to 50μm isotropic resolution and converted the MRI labels into an image Nissl-like contrast using the trained GAN model. Nissl-contrast images were re-stacked into a 3D volume to act as an anatomical prior for registration.
We performed an iterative affine registration procedure between the MRI-based shape prior and the 3D stack of histological sections. This process was repeated for a total 5 iterations, producing a final 3D volume with aligned coronal slices and a global shape approximately matched to the in utero fetal brain
see: align_2D_to_3D_reference.py
To create the final 3D volume, we employed a data augmentation technique, generating n=50 unique representations of the affinely-aligned data by applying nonlinear distortions along all three image axes. For each volume, we performed a weighted nonlinear registration between neighbouring sections to account for residual misalignments. Finally, to create a smooth 3D reconstructed volume, we co-registered all 50 augmented and aligned volumes into a single probabilistic anatomical template with voxel resolution 150 × 150 × 150μm using an iterative, whole-brain nonlinear registration
see: create_template.py
To reconstruct the fetal cortical surface, we adapted existing protocols for ex vivo and non-human primate surface reconstruction with Freesurfer. We used the μBrain tissue labels to generate a ‘white matter’ mask (all subcortical structures and tissue zones, excluding the cortical plate). We used this mask to generate inner and outer surfaces for the μBrain volume.
see: run_freesurfer_commands_1.sh
see: run_freesurfer_commands_2.sh
We aligned the μBrain cortical surface to the earliest timepoint of the dHCP fetal template surface using a two-step nonlinear surface registration guided by a set of anatomical priors. We used MSM to perform an initial nonlinear spherical registration between μBrain and dHCP surfaces based on alignment of sulcal depth. After this, we created a set of coarse cortical labels on the dHCP surface matched to corresponding μBrain labels by combining a) dHCP cortical atlas labels,87 b) manual labels guided by sulcal anatomy on the 36 week fetal surface and c) combining μBrain labels in the same lobes into single anatomical labels. A secondary multivariate spherical registration between μBrain and fetal surfaces was initialised using the previously calculated sulcal alignment and driven by alignment of cortical ROIs across surfaces
We used publicly-available microarray data from four prenatal brain specimens aged 15 to 21 PCW. After updating gene assignments using Reannotator, see details here, we removed any probes assigned to more than one gene and low signal probes designated ‘absent’ (34.67% of probes). We also removed tissue samples from the marginal zone, subpial granular zone and subcortical and midbrain structures (54.46% of samples). Where multiple probes mapped to a single gene, the probe with the highest differential stability (DS) was selected.
Tissue sample loctions were matched to cortical regions in the μBrain atlas (all_structure_lut.csv) and, where more than one sample was available for a given region or zone, e.g.: samples from the outer and inner cortical plate in the same region, gene expression was averaged across samples. Finally, any probes with missing data in more than 10% of tissue samples were removed (n=1253). This resulted in expression data from 8771 genes from up to 27 regions and 5 tissue zones:
see: process_microarray_data.py
- remove vertex outliers from cortical area data using a sliding window approach
- model total cortical surface area as a function of age
- fit vertexwise log-log models to cortical area
- output areal scaling maps and scaling for each μBrain region
see: 02-gene-correlations.ipynb
- perform PCA on preprocessed microarray data
- perform ANOVA for main effects of Zone, Region and Timepoint to identify ZRT genes
see: 03-ZRT-gene-analysis.ipynb
- test age effects on ZRT gene expression and identify up- and down-regulated ZRT genes in each tissue zone
- enrichment of specific cell type markers in ZRT genes
- average expression of ZRT genes at each timepoint across all zones and regions
see: 04-gene-expression-and-areal-scaling.ipynb
- perform correlations between regional allometric scaling and gene expression
- identify tissue location of significant associations
see: 05-scaling-characterisation.ipynb
- cortical cell type specificity of ZRT scaling genes
- expression of hypo- and hyper-allometric ZRTscaling and ZRTneo genes
- cell type enrichment of ZRTscaling and ZRTneo genes
- expression of MDK and CUX1 over all regions and tissue zones
- enrichment of ZRT in genes located near to open chromatin regions of the prenatal brain
- laminar specific enrichment
- identification of TF motifs
Ding, S.-L. et al. Cellular resolution anatomical and molecular atlases for prenatal human brains. J. Comp. Neurol. 530, 6–503 (2022).
Miller, J. A. et al. Transcriptional landscape of the prenatal human brain. Nature 508, 199–206 (2014).
Majka, P. & Wójcik, D. K. Possum—A Framework for Three-Dimensional Reconstruction of Brain Images from Serial Sections. Neuroinformatics 14, 265–278 (2016).
Isola, P., Zhu, J.-Y., Zhou, T. & Efros, A. A. Image-to-Image Translation with Conditional Adversarial Networks. ArXiv161107004 Cs (2018).
Gholipour, A. et al. A normative spatiotemporal MRI atlas of the fetal brain for automatic segmentation and analysis of early brain growth. Sci. Rep. 7, 476 (2017).
Karolis, V. et al. Developing Human Connectome Project spatio-temporal surface atlas of the fetal brain. 300 MiB (2023) doi:10.12751/G-NODE.QJ5HS7.
Robinson, E. C. et al. Multimodal surface matching with higher-order smoothness constraints. NeuroImage 167, 453–465 (2018).