Graduate Student University of Virginia Charlottesville, Virginia
Abstract Text: Introduction
Cortical parcellation is an indispensable tool for studying the structure, function, and organization of the healthy and diseased brain and for identifying potential targets for therapeutic interventions. Large-scale cortical parcellation relies on the diffeomorphic registration of a subject’s brain to a standardized atlas template with labeled cortical regions. The structural T1-weighted scans used for this purpose have excellent gray-white matter contrast but no sensitivity to cortical cytoarchitectonic features. Consequently, atlas-registration-based parcellations can accurately delineate boundaries between cortical areas aligned with gross anatomical landmarks (sulci, gyri), but their lack of cytoarchitectonic contrast makes a finer parcellation problematic.
High-resolution mean apparent propagator (MAP) MRI has high sensitivity to cortical cytoarchitecture revealing areal boundaries and lamination patterns. We assess the feasibility of using local k-means clustering of voxelwise high-resolution MAP parameters to refine the cortical parcellation estimation obtained with atlas-based registration and to generate subject-specific segmentations of cytoarchitectonic domains observed with histology.
Methods
We scanned a perfusion-fixed macaque monkey brain using a MAP-MRI protocol with 200µm spatial resolution, TE/TR=50/650ms, 112 DWIs with multiple b-values and orientations. After post-processing, we estimated the MAP coefficients and computed the microstructural parameters: propagator anisotropy (PA), return-to-axis probability (RTAP), and non-gaussianity (NG). In the same session, we acquired structural scans from which we segmented the gray-white matter boundary. After imaging, we sectioned the brain into 50µm-thick coronal slices which we processed with multiple histological stains.
We registered the MAP-MRI data to the D99 digital macaque brain atlas and partitioned the cortical labels into 11 major contiguous regions R1-11, in each hemisphere, that roughly correspond to parts of the major lobes (e.g., prefrontal, temporal, occipital, etc.) and are separated by distinctive anatomical landmarks like the major sulci and gyri that can be robustly delineated with atlas-registration-based parcellation. Within each region R1-11, we performed k-means clustering of all voxels using the MAP-MRI parameters PA, RTAP, and NG, and the distance from the GM/WM boundary as features. We quantified the correspondence between the D99 (atlas-based) and MAP (cytoarchitectonic) segmentations within each region, compared their topologies, cross-tabulated and matched their labels, and assessed their accuracies by comparing them with the corresponding histological images.
Results
The topologies of the 157 D99 cortical labels grouped into 11 major brain regions were consistent in both hemispheres. The MAP-based cytoarchitectonic segmentation of the 3D cortex data showed distinct laminar and areal boundaries with a high degree of symmetry between hemispheres. The segmentation labels were matched in both hemispheres based on their median MAP-parameter values.
The MAP-derived cytoarchitectonic domains show cortical laminar patterns that are not present in the D99 atlas. Discontinuities between these patterns can be matched to D99 cortical labels, although the boundaries estimated with the MAP-based algorithm are more consistent with histological images. Some MAP cytoarchitectonic domains correspond to different layers that extend across areal boundaries, while others terminate abruptly. The median MAP parameter values within each domain represent the k-centroid coordinates and reveal the main differences in microstructural features compared to neighboring domains. Median values of PA, NG, and RTAP computed along the borders of the 11 major regions showed high contrast.
Quantifying the cross-tabulation, a.k.a. the contingency matrix between the MAP and D99 parcellations, we found that 90% of the volume of any D99 cortical label was covered by at most five distinct MAP labels, potentially reflecting the presence of a laminar pattern. Meanwhile, 99% of the volume of all 157 D99 cortical labels was completely covered using only 367 MAP labels.
Discussion
Our study enables a fine segmentation of cytoarchitectonic domains by leveraging the microstructural sensitivity of high-resolution MAP parameters and by focusing separately on 11 contiguous cortical subregions. K-means clustering of voxelwise MAP parameters within these subregions delineates laminar patterns and identifies boundaries between cortical areas in better agreement with histology than conventional atlas-based cortical parcellation. These findings corroborate the excellent microstructural sensitivity of MAP-MRI and support its potential for direct segmentation of healthy and injured cytoarchitectonic domains. It could complement histological analysis, improve current parcellation methods, and enable the construction of digital 3D cytoarchitectonic brain atlases for use in clinical and neuroscience research.
