The Limitation of Volumetric Brain Analysis
Traditional structural neuroimaging relies heavily on voxel-based morphometry, which treats the brain as a rigid three-dimensional grid of pixels (voxels). This approach measures the density or total volume of gray matter within specific regions. However, the human cerebral cortex is biologically a highly folded, continuous two-dimensional sheet of tissue—a 2D manifold embedded in 3D space.
Measuring raw volume conflates two entirely distinct morphometric properties: cortical thickness and surface area. Because volume is roughly the product of thickness and surface area, a reduction in a region's volume could mean the cortical ribbon is thinning, or it could mean the surface area is shrinking due to altered folding patterns.
Voxel-Based (3D Grid) Surface-Based (2D Manifold Mesh)
┌───┬───┬───┐ ▲ Outer Pial Surface
│ █ │ █ │ █ │ ├─► ──────────────────────
├───┼───┼───┤ │ │ Cortical Thickness
│ █ │ █ │ █ │ │ ▼ (Euclidean Distance)
└───┴───┴───┘ └─► ──────────────────────
Conflates thickness & area Inner White Matter Surface
These two metrics reflect completely different neurobiological processes. Cortical thickness is primarily determined by the vertical laminar architecture of cortical columns (the number of neurons per column), whereas surface area is driven by the horizontal proliferation of these columns and the mechanical forces that drive brain folding during early development. To untangle these signals, researchers use Surface-Based Morphometry (SBM) to reconstruct and analyze the boundaries of the cortex explicitly.
Reconstructing the Cortical Manifold
To isolate thickness and surface area, SBM algorithms transform 3D MRI voxels into geometric surface meshes. The pipeline identifies two critical boundaries:
- The inner white matter surface, where gray matter meets white matter.
- The outer pial surface, where gray matter meets cerebrospinal fluid.
As detailed by Ruthger Righart et al. in Volume versus surface-based cortical thickness measurements: A comparative study with healthy controls and multiple sclerosis patients, these boundaries are modeled as continuous triangular meshes consisting of tens of thousands of vertices. Once these meshes are aligned, cortical thickness is calculated by measuring the Euclidean distance between corresponding vertices on the inner and outer surfaces.
This boundary-distance modeling approach, also evaluated longitudinally by Benita Schmitz‐Koep et al. in Longitudinal assessment of cortical thickness in healthy older individuals: a comparison between CAT12 and freesurfer, provides a vertex-by-vertex map of the cortical ribbon. Because measurements follow the geometry of the tissue rather than an arbitrary 3D grid, SBM achieves high anatomical specificity and remains invariant to global variations in brain size or orientation.
Disentangling Thickness from Surface Area
Separating thickness from surface area reveals structural signatures that volume-based methods obscure. Because surface area expands dramatically during late fetal development while thickness is dynamically modified by synaptic pruning during adolescence and dendritic atrophy in aging, the two metrics serve as distinct biomarkers for development and disease.
The authors of Joint Analysis of Cortical Area and Thickness as a Replacement for the Analysis of the Volume of the Cerebral Cortex argue that analyzing area and thickness jointly provides far greater statistical sensitivity than traditional volumetric analysis. For instance, in psychiatric conditions like Major Depressive Disorder, Ting Zheng demonstrated in Sex differences of brain cortical structure in major depressive disorder that while cortical thickness alterations were subtle and failed to survive strict statistical correction, cortical surface area served as a highly robust marker for identifying sex-specific structural changes.
Conversely, in neurodegenerative pathologies like Alzheimer's disease, tracking localized thinning over time provides a direct readout of progressive cellular loss. For example, Elijah Mak et al. used surface reconstructions in Longitudinal associations between β-amyloid and cortical thickness in mild cognitive impairment to demonstrate that amyloid-beta accumulation in specific regions accelerates cortical thinning in downstream, non-local areas, mapping the precise spatial trajectory of neurodegeneration.
Quantifying the Wrinkles: The Local Gyrification Index
The human cortex is defined by its complex folds—the ridges (gyri) and valleys (sulci). Standard thickness and area metrics fail to capture the spatial complexity of this folding pattern. To quantify this geometry, SBM introduces the Local Gyrification Index (LGI).
In Surface Vulnerability of Cerebral Cortex to Major Depressive Disorder, Feng Shi explains that the LGI measures the ratio of the total cortical surface area (including the extensive tissue hidden deep within the sulcal folds) to the visible, exposed outer surface area.
$$\text{LGI} = \frac{\text{Total Surface Area (In Sulci + Gyri)}}{\text{Visible Outer Surface Area}}$$
A highly folded region will have a high LGI, whereas a smooth, atrophied region will have an LGI approaching 1. Because cortical folding patterns are largely locked in during the second trimester of pregnancy and remain stable throughout life, LGI serves as a proxy for identifying neurodevelopmental structural signatures. This allowed Colleen Rollins et al. in Evidence in cortical folding patterns for prenatal predispositions to hallucinations in schizophrenia to trace specific sulcal deviations back to prenatal predispositions.
Furthermore, as the brain undergoes neurodegeneration, these folds alter. Tao Liu et al. demonstrated in Cortical Gyrification and Sulcal Spans in Early Stage Alzheimer's Disease that geometric measures of brain folding, such as the global sulcal index and individual sulcal spans (the 3D distance between opposing banks of a fold), offer greater statistical power for distinguishing early stages of cognitive impairment than thickness measurements alone. As the cortex atrophies, the gaps between folds widen and the overall structure becomes measurably smoother.
Geometric Deep Learning on Cortical Meshes
Representing the brain as a surface mesh opens the door to advanced geometric machine learning. Traditional convolutional neural networks require rigid 3D voxel grids, which struggle with the highly irregular, curved geometry of the cortex. By treating the cortical mesh as a graph where vertices are nodes and mesh connections are edges, researchers can apply Graph Convolutional Neural Networks (gCNNs) directly to the manifold.
The authors of Geometric deep learning on brain shape predicts sex and age demonstrated that the geometric folding patterns extracted from these meshes contain highly predictive structural signatures, allowing gCNNs to accurately predict demographic traits like age and sex from topology alone, independent of tissue thickness.
Beyond the cortex, this shape-based framework is being applied to subcortical structures. In S‐GMAS: Genome‐Wide Mediation Analysis With Brain Subcortical Shape Mediators, Li Shen showed that modeling subcortical structures using advanced functional shape representations uncovers significant genetic mediation effects on cognition that traditional, simple volumetric measures completely fail to detect. By treating brain structures as continuous geometric surfaces rather than static volumes, surface-based morphometry captures the nuanced mathematical signatures of human development, genetics, and disease.