Mapping the individual human cortex using multidimensional MRI and unsupervised learning

Abstract

Human evolution has seen the development of higher-order cognitive and social capabilities in conjunction with the unique laminar cytoarchitecture of the human cortex. Moreover, early-life cortical maldevelopment has been associated with various neurodevelopmental diseases. Despite these connections, there is currently no noninvasive technique available for imaging the detailed cortical laminar structure. This study aims to address this scientific and clinical gap by introducing an approach for imaging human cortical lamina. This method combines diffusion-relaxation multidimensional MRI with a tailored unsupervised machine learning approach that introduces enhanced microstructural sensitivity. This new imaging method simultaneously encodes the microstructure, the local chemical composition and importantly their correlation within complex and heterogenous tissue. To validate our approach, we compared the intra-cortical layers obtained using our ex vivo MRI-based method with those derived from Nissl staining of postmortem human brain specimens. The integration of unsupervised learning with diffusion-relaxation correlation MRI generated maps that demonstrate sensitivity to areal differences in cytoarchitectonic features observed in histology. Significantly, our observations revealed layer-specific diffusion-relaxation signatures, showing reductions in both relaxation times and diffusivities at the deeper cortical levels. These findings suggest a radial decrease in myelin content and changes in cell size and anisotropy, reflecting variations in both cytoarchitecture and myeloarchitecture. Additionally, we demonstrated that 1D relaxation and high-order diffusion MRI scalar indices, even when aggregated and used jointly in a multimodal fashion, cannot disentangle the cortical layers. Looking ahead, our technique holds the potential to open new avenues of research in human neurodevelopment and the vast array of disorders caused by disruptions in neurodevelopment.

Keywords: cortical parcellation; diffusion–relaxation; machine learning; microstructure imaging; multidimensional MRI.

Graphical Abstract

Figure 1

System diagram. Voxelwise diffusion–relaxation probability distributions in the image domain were transformed to the transport domain, where each spectrum is represented as a point on a high-dimensional Riemannian manifold. The shortest distance between a point on the manifold and the mean distribution I₀ can be computed using the modified two-Wasserstein distance. The linear optimal transport (LOT) embedding provides a linearized version of this distance and enables geodesic distances to be approximated as Euclidean distances in the transport domain. Unsupervised learning was performed in the transport domain using the diffusion–relaxation distributions alone, without their spatial coordinates, and the cluster assignments were visualized in the image domain as segmented cortical images. Note that when k = 3 clusters are used, WM and two GM components are distinguished.

Figure 2

Histological findings from the three examined cases. Cellular morphology is demonstrated in representative Nissl-stained sections from Subjects 1–3 in A–C, respectively. Entire cortical cross-sections are shown in the top panel and magnified regions in the bottom panel. Delineated cortical layers were overlayed on the magnified regions from each subject, depicting region-specific laminar variations in the density of neuronal cells and their architectural patterns.

Figure 3

Intra-cortical diffusion–relaxation signature. A transcortical region of interest from Subject 1, and single voxel depth-specific T₁–T₂, T₁–MD and T₂–MD spectra are shown with logarithmic scale from the subcortical WM margin out to the cortical surface. The main diffusion peak shifted from 0.85 to 0.54 µm²/ms as a function of depth, along with the appearance of a slow diffusion component that intensifies as a function of depth from the cortical surface. In addition, the main T₁ and T₂ components shifted from 339.3 and 57.8 ms at the cortical surface to 281.2 and 30.7 ms, respectively, at the cortical depth. Finally, a clear short T₁ component appears at about 1 mm into the cortex and intensifies as a function of depth. To facilitate visualization of the logarithmic scale, an asterisk symbol marks the location of the main signal components at the cortical surface. MD, mean diffusivity.

Figure 4

Segmentation spatial contiguity and comparison of 1D and multidimensional MRI. Segmented cortical volumes from Subjects 1 to 3 are shown in A–C, respectively. A representative slice (x-y plane) is shown for each subject, along with four horizontal and four vertical cross-sections in the z direction. The top and bottom panels show the multidimensional spectra- and scalar-based clustering results. Diffusion–relaxation MRI data result in five distinct and spatially consistent clusters, labelled GM 1–GM 5 from the outermost to innermost cortical layer, respectively. Layered and contiguous laminar structures were demonstrated across the volume in all samples. Conversely, scalar MRI data result in roughly three consistently distinguishable regions, and the cluster maps appear noisy and spatially inconsistent. GM, grey matter.

Figure 5

Multimodal comparison of MRI- and histology-based cortical parcellation. Representative slices of MRI (top)- and histology (bottom)-based cortical lamina segmentation from Subjects 1 to 3 are shown in A–C, respectively. In all cases, the MRI-based maps enabled delineation of contiguous cortical lamina based on T₁–T₂–MD spectral signatures, although our clustering approach was agnostic to location in space. Qualitative agreement is evident from visual inspection. Limitations in inter-modality co-registration of MRI and histological data sets hinder optimal image alignment, which may result in slice plane differences, leading to discrepancy in the apparent layer thicknesses.

Figure 6

Layer-specific multidimensional MRI signatures. The mean distributions from each cluster, averaged across the entire sample volume from all three subjects in the image domain, are shown for (A) T₁–T₂, (B) T₁–MD and (C) T₂–MD. The average pairwise LOT distance between each cluster centroid is summarized in distance matrices, shown in the rightmost column. A gradual shift towards shorter T₁ and T₂ values, as well as the emergence of slow diffusivity and short T₁ components, is evident from the outermost GM layer to WM. From the distance matrices, it is evident that radially contiguous clusters have centroids that are more similar in the LOT metric space, and that T₁–T₂ and T₂–MD yield consistently higher inter-layer distances compared with T₁–MD. LOT, linear optimal transport; MD, mean diffusivity; WM, white matter; GM, grey matter.