Smarter Brain Imaging: Anatomically Constrained Smoothing Boosts fMRI Reliability While Dodging Artifacts

For decades, neuroimaging researchers have relied on a seemingly innocuous step to sharpen their brain scans: spatial smoothing. By blurring fMRI data across neighboring voxels, scientists boost the signal-to-noise ratio and increase the statistical power needed to detect brain activation. But a new study published in Frontiers in Neuroimaging suggests this fundamental technique may be quietly distorting the very brain maps it is meant to clarify. David G. Ellis and Michele R. Aizenberg from the University of Nebraska Medical Center tested the effects of unconstrained Gaussian smoothing against an anatomically constrained volumetric smoothing method that prevents signal from bleeding across the white and gray matter surfaces of the cortex. Their findings reveal that while smoothing does enhance reliability, unconstrained smoothing spreads activation across cortical boundaries, inflates white matter activation, and biases functional connectivity metrics—artifacts that constrained smoothing largely avoids.

The core problem with conventional smoothing lies in its indifference to brain anatomy. A Gaussian kernel treats the brain as a continuous 3D volume, aggregating signal from any voxel within its radius regardless of whether those voxels belong to the same gyrus, the same cortical layer, or even the same tissue type. When the cortex folds in complex sulcal patterns, two voxels that sit just millimeters apart in Euclidean space may be separated by a deep sulcus and belong to functionally distinct regions. Smoothing across that boundary blurs signal from one gyrus into its neighbor, creating activation clusters that never existed biologically. The white-gray matter interface presents an even more vulnerable boundary, as signal from cortical gray matter can leak into white matter and artificially shift the apparent peak of activation away from the cortex.

Surface-based fMRI analysis has long been considered the gold standard for avoiding these volumetric smoothing artifacts. By projecting data onto an unfolded cortical surface, this approach inherently respects the geodesic distance along the cortex and prevents signal from crossing sulci. But surface-based workflows require complex preprocessing pipelines, specialized file formats like CIFTI, and a small set of tools that may not integrate with researchers’ preferred software. Many established neuroimaging packages still lack surface-based capabilities, leaving most labs to choose between adopting an entirely new analytical paradigm or continuing with volumetric smoothing and accepting its artifacts.

Ellis and Aizenberg built their constrained smoothing tool as a practical middle ground. The method represents each brain voxel as a node in a graph, with edges connecting all 26 neighboring voxels in 3D space. Critically, the researchers then prune any edge that crosses either the pial or white matter surface as reconstructed by FreeSurfer, using ray tracing to rapidly evaluate surface intersections across the more than 40 million nodes in a typical brain graph. After pruning, the graph separates into four independently smoothable components—left and right gray matter and left and right white matter—ensuring that smoothing kernels can never aggregate signal across the cortical boundaries. The researchers implemented the approach in Python, avoiding the licensing costs of MATLAB, and optimized the heat kernel smoothing parameter to approximate a user-specified Gaussian full-width at half-maximum, allowing direct comparison between constrained and unconstrained smoothing at equivalent kernel widths.

To evaluate the method, the researchers applied both smoothing approaches to simulated data, a sensory task fMRI dataset, a precision motor mapping dataset from the Midnight Scan Club, and resting state fMRI data from the Human Connectome Project in Aging. In the simulation experiments, where 68 cortical regions were isolated and white noise added, constrained smoothing accurately recovered the shape of the original regions while Gaussian smoothing created artifacts by smoothing across boundaries. Constrained smoothing demonstrated higher accuracy in terms of both Dice similarity coefficient and Pearson correlation across all tested kernel widths, and more accurately concentrated positive predictions within gray matter while reducing white matter activation.

The real-world task fMRI results told a similarly compelling story. In the sensory task analysis, where participants received brushing stimulation to the left hand, both Gaussian and constrained smoothing increased the reliability of activations across repeated acquisitions. However, Gaussian smoothing significantly increased the percentage of active voxels in white matter relative to constrained smoothing. In the motor task experiments, where each of ten subjects completed 20 runs of finger tapping, toe tapping, and tongue movement, Gaussian smoothing with kernel widths greater than 3 mm decreased the accuracy of individual session maps compared to constrained smoothing. Notably, smoothing did not uniformly improve accuracy across all metrics—Dice scores were actually highest with no smoothing at all, suggesting that the reliability gains from smoothing come at a cost to spatial precision.

The false positive analysis addressed a critical statistical concern. Because constrained smoothing is inherently anisotropic, it produces smoothed data that may violate the Gaussian random field assumptions underlying parametric statistical models in packages like FSL FEAT. Previous research by Eklund and colleagues has shown that such violations can lead to inflated false positive rates in fMRI studies. Ellis and Aizenberg tested this by running task fMRI analysis on resting state data—data acquired with no task being performed—so that any detected activation could be considered a false positive. With cluster significance thresholding, mean false positive voxel percentages remained below 5% for both smoothing methods across all tested kernel widths, providing reassurance that constrained smoothing does not dramatically inflate statistical errors when appropriate thresholding is applied.

Perhaps the most surprising findings emerged from the resting state connectivity analysis. Using the Schaefer/Yeo 1,000-parcel atlas, the researchers fit generalized additive mixed models to examine how smoothing altered the relationship between Euclidean distance and functional connectivity. Both Gaussian and constrained smoothing inflated connectivity between spatially proximal regions, but Gaussian smoothing demonstrated more inflation at kernel widths of 6 mm or greater. Even more striking, both smoothing methods significantly altered graph theory metrics including clustering coefficient, global efficiency, local efficiency, and modularity under all tested kernel widths. Constrained smoothing had less effect on local efficiency, global efficiency, and clustering coefficient at higher kernel widths, though the differences between methods were marginal compared to the overall effect of smoothing itself on these network measures.

The researchers acknowledge an important caveat: constrained smoothing is inherently more conservative than unconstrained smoothing at identical nominal kernel widths. By restricting connections across cortical boundaries, constrained smoothing aggregates fewer data points, effectively lowering the dimensionality of the smoothing operation. The effective level of smoothing also depends on graph connectivity, which in turn depends on voxel size—smaller voxels produce more connections and more data aggregation. This suggests constrained smoothing may be most advantageous in high-resolution fMRI experiments, where the additional anatomical detail justifies the computational cost of graph construction and pruning.

As a practical alternative to both unconstrained volumetric smoothing and full surface-based analysis, anatomically constrained volumetric smoothing offers researchers a way to enhance fMRI reliability while keeping data in the familiar voxel space. The method requires reconstructed cortical surfaces but avoids the need to transform data into surface-based representations, preserving compatibility with decades of established volumetric processing pipelines. Future work could extend this graph-based approach to smooth along white matter tracts, integrate with Bayesian inference frameworks that determine optimal smoothing from the data, or apply nonparametric permutation testing for valid statistical inference on constrained-smoothed volumetric data. For now, the study provides compelling evidence that the smoothing step so many researchers take for granted deserves a second look—and that respecting the brain’s own boundaries may be the key to more accurate brain maps.

Subject of Research: Anatomically constrained volumetric smoothing enhances fMRI reliability while avoiding smoothing artifacts

Article Title: Anatomically constrained volumetric smoothing enhances fMRI reliability while avoiding smoothing artifacts

Article References: Ellis, D. G., & Aizenberg, M. R. (2026). Anatomically constrained volumetric smoothing enhances fMRI reliability while avoiding smoothing artifacts. Frontiers in Neuroimaging, 5, Article 1753534. https://doi.org/10.3389/fnimg.2026.1753534

Image Credits: AI Generated

DOI: 10.3389/fnimg.2026.1753534

Keywords: Anatomically, constrained, volumetric, smoothing, enhances, fMRI, reliability, while, avoiding, artifacts, scientific research