Geometry of the cumulant series in diffusion MRI

Seeing Hidden Structure in the Brain

Magnetic resonance imaging is already a powerful window into the living brain, but standard scans mostly show anatomy. This work explains how a special type of MRI, called diffusion MRI, can be reinterpreted using ideas from geometry and symmetry to reveal much finer details of brain tissue. By treating the signal as a combination of simple building blocks with well-defined symmetries, the authors show how to extract compact, hardware-independent fingerprints of microstructure that can aid diagnosis and make advanced scans faster and more practical.

How Water Motion Reveals Brain Microstructure

Diffusion MRI tracks how water molecules wiggle over microscopic distances inside each image voxel. In brain tissue, water motion is restricted by cells, fibers, and membranes, so the way it diffuses carries information about the underlying microstructure. For years, most clinical scans have focused on a single quantity called the diffusion tensor, which treats water motion as roughly Gaussian and summarizes it with a 3×3 matrix. This yields familiar measures such as mean diffusivity and fractional anisotropy, widely used to map white matter pathways. However, the actual signal is richer: deviations from simple Gaussian behavior contain clues about tissue heterogeneity, cell shape, and more. The article tackles the question of how much information is really present in this signal and how best to organize it.

From Complex Tensors to Simple Invariants

The authors describe the diffusion signal using a mathematical expansion in terms of “cumulants,” which are higher-order summaries of how water displacements deviate from a simple bell-shaped distribution. Each cumulant is a tensor, an object with many components that change when you rotate the coordinate system. Instead of working with these raw components, the team uses the rotational symmetry of three-dimensional space to break each tensor into irreducible pieces that transform in simple, predictable ways under rotation. From these pieces they build scalar quantities called invariants, which have the same value no matter how the head is oriented in the scanner. This procedure, guided by group theory, reveals that up to the second order in diffusion weighting, the key information in the signal can be captured by 3 invariants from the basic diffusion tensor and 18 from the next-order covariance tensor, together forming what they call the RICE description.

Linking Geometry to Tissue Properties

Crucially, the invariants are not just abstract numbers. They have clear geometric and physical interpretations in terms of “size” and “shape” variations of microscopic diffusion ellipsoids within each voxel. Some invariants describe how much diffusivities vary from one tiny compartment to another, while others describe how those variations are oriented relative to each other. Well-known diffusion MRI metrics such as mean diffusivity, mean kurtosis, microscopic fractional anisotropy, and measures of isotropic and anisotropic variance all turn out to be specific combinations of only seven of these invariants. The remaining 14 from the covariance tensor, as well as additional invariants related to truly non-Gaussian water motion, constitute largely unexplored contrasts that may be sensitive to subtle changes in microstructure, such as fiber crossings or shifts in cell shape.

Putting the Method to the Test in Disease

To assess clinical relevance, the authors applied their framework to a large set of real-world brain scans from 1189 people, including 627 patients with multiple sclerosis and 562 matched controls. These clinical exams used standard diffusion MRI protocols that only sample a limited part of the full tensor space. Even under this restriction, the researchers could compute all invariants associated with the conventional kurtosis tensor. When they used these invariants as inputs to simple logistic regression models, they found systematically better classification of multiple sclerosis than when they used only traditional diffusion and kurtosis metrics. In some white matter regions, the error in ranking patients versus controls decreased by up to 30 percent, without acquiring any extra data, purely by reorganizing the existing signal through symmetry-based invariants.

Designing Faster, More Efficient Scans

Another practical payoff of the geometric viewpoint is scan optimization. By exploiting the connection between tensor symmetries and the way measurements are distributed on spheres of diffusion directions, the authors designed minimal acquisition schemes that still allow unbiased estimation of the most commonly used invariants. Using clever arrangements of only six diffusion directions per shell, based on the vertices of simple geometric shapes, they show that key maps such as mean diffusivity, fractional anisotropy, mean kurtosis, and microscopic fractional anisotropy can be obtained in about one to two minutes for the whole brain. These “instant RICE” protocols dramatically shorten scan time compared with conventional approaches while preserving the essential information content.

Why This Matters for Future Brain Imaging

Overall, the study shows that diffusion MRI signals can be reorganized into a compact set of rotation-invariant numbers that reflect distinct geometric aspects of tissue microstructure. Many of these invariants have not yet been explored in biology, but the initial results in multiple sclerosis suggest they contain clinically useful information. Because they are defined independently of scanner hardware and head orientation, these scalar maps are natural candidates for feeding into machine learning systems aimed at detecting disease, tracking development, or studying aging across large populations. At the same time, the proposed fast protocols promise to bring more advanced diffusion contrasts into routine clinical practice without prohibitive scan times.

Citation: Coelho, S., Chen, J., Szczepankiewicz, F. et al. Geometry of the cumulant series in diffusion MRI. Nat Commun 17, 4220 (2026). https://doi.org/10.1038/s41467-026-70018-w

Keywords: diffusion MRI, brain microstructure, tensor invariants, multiple sclerosis, medical imaging