A central claim in early Alzheimer's research is that brain function breaks down before brain structure does. In people who are still cognitively normal but carry abnormal spinal-fluid biomarkers, resting-state fMRI shows weaker local and network coherence, while structural MRI still looks normal. This note argues that part of that "function first" signal may be a geometric artifact. Many fMRI coherence measures compare each voxel with its spatial neighbors. In the thin, folded cortical ribbon, a voxel near the edge of gray matter, next to cerebrospinal fluid, white matter, or a parcel border, has fewer valid neighbors. Subtle atrophy that is too small to flag as structural loss still moves more voxels into this edge zone. A coherence statistic that ignores truncated neighborhoods will then report lower coherence even if the underlying signal has not changed.
We propose a coherence measure that is calibrated against each subject's own gray-matter geometry. We then use it to ask, in ADNI's amyloid-positive cognitively normal participants, how much of the reported functional–structural dissociation survives once boundary exposure is accounted for. Either outcome is informative. If the dissociation survives, the "function precedes structure" claim gets a stronger footing. If it shrinks, a widely used class of early biomarkers is partly measuring geometry.
Core Claim
Work on finite wireless sensor networks shows that nodes near a domain's edge lose about half of their interaction neighborhood, and corner nodes up to three quarters. That loss, not anything intrinsic to the nodes, drives coverage holes and early disconnection (Results on Finite Wireless Sensor Networks: Connectivity and Coverage). Analyses that ignore borders systematically misstate connectivity requirements and reliability (Connectivity in Secure Wireless Sensor Networks under Transmission Constraints; Performance Analysis of Connectivity and Localization in Multi-Hop Underwater Optical Wireless Sensor Networks).
The cortex at fMRI resolution is an extreme version of this setting. With a ribbon 2–4 mm thick sampled at 2–3 mm voxels, almost every gray-matter voxel is a boundary voxel. Our claim is that local spatial autocorrelation of BOLD signals (local Moran's I, a voxel-level score of how much a location's signal resembles its neighbors' signals) depends on how much of each voxel's neighborhood is truncated. Preclinical atrophy changes that truncation. The result is a group difference that looks functional but is driven by structure.
For a voxel $i$, define boundary exposure as
$$e_i = 1 - \frac{|N_r(i) \cap \Omega|}{|N_r(i)|},$$
where $N_r(i)$ is the full ball of radius $r$ around $i$ and $\Omega$ is the subject's gray-matter (or parcel) mask. The corrected statistic is
$$z_i = \frac{I_i - \mathbb{E}_0[I_i \mid \Omega]}{\mathrm{SD}_0[I_i \mid \Omega]}.$$
Here $I_i$ is the observed local Moran's I. The expectation and standard deviation are computed under null surrogate fields with a fixed spatial autocorrelation length, simulated on that subject's own mask $\Omega$. Because the null is built on the exact finite domain, it removes whatever bias the geometry induces, without assuming any particular functional form for that bias.
Research Program
The approach is shaped by a constraint from the theory of fields on graphs. Parametric estimation for Gaussian fields indexed by graphs derives an explicit boundary-correction matrix for autoregressive fields observed on a finite piece of a graph. Its guarantees require amenability, meaning the boundary must shrink relative to volume as the observed region grows. The cortical ribbon violates this at fixed resolution, because its boundary-to-volume ratio is of order one. So we do not try to port an asymptotic correction. We simulate on the actual finite domain instead.
The torus results point the same way. Cube Versus Torus Models and the Euclidean Minimum Spanning Tree Constant shows that edges stop mattering only as density grows without limit. Spectral analysis of virus spreading in random geometric networks and Sharp exact recovery threshold for two-community Euclidean random graphs show that removing edges distorts thresholds in finite graphs. Neither periodic wrapping nor infinite-domain formulas are safe here.
The program has three parts.
Part 1: geometry-matched null maps. For each ADNI subject with resting-state fMRI, T1 MRI and CSF biomarkers, we segment gray matter. We follow Eslami et al.'s interior/edge/corner zoning, carried over to voxels, and compute $e_i$ at several radii and for two kinds of boundary: the tissue edge and the atlas parcel edge (AAL116 and Yeo networks, matching Ada-FCN and mBrainGT). We then generate a few hundred surrogate BOLD fields per subject, with autocorrelation lengths spanning the range estimated from that subject's own residuals. From these we form the null moments in the equation above.
Part 2: phantom-effect estimation. We compare amyloid-positive cognitively normal participants with amyloid-negative ones (CSF tTau/Aβ42 defined as in Wei et al.) on surrogate data alone. Any group difference in uncorrected $I_i$ under the null is, by construction, a pure geometry effect. We express its size relative to the observed effect.
Part 3: re-analysis of the real data with corrected maps. We rerun the real-data group comparison on $z_i$. We also extend the correction to correlations between adjacent parcels, where misregistration and partial-volume mixing near parcel borders in atrophied brains can inflate or deflate connectivity.
A secondary target is the unresolved disagreement over the default mode network (DMN). Functional Connectome Signatures of Patients with Asymptomatic and Typical Alzheimer's and Mapping Structural Connectome Patterns in Alzheimer's Disease place early vulnerability in the DMN. Wei et al. instead find that cognitive-control and visual networks dominate detection, with the DMN less critical in the asymptomatic phase. The DMN's medial parietal hubs sit next to the medial wall and are prone to atrophy, so they have unusually high boundary exposure. We test whether network rankings change after correction, and whether the size of each network's change tracks its mean exposure. A positive answer would tie the contradiction to measurement geometry rather than biology.
Baselines
The naive reference is uncorrected local Moran's I computed on volume data with standard preprocessing. Any functional group difference reported here is what the field would currently accept. Every corrected variant must be compared against it on identical subjects, preprocessing and atlas.
The strongest simple competitor is covariate regression: adding regional cortical thickness and gray-matter volume as nuisance covariates in the group model. Every structural–functional study can do this cheaply. It gets the same subjects and the same structural measures our exposure map uses. The geometry-matched null beats it only if, on surrogate data where the true functional effect is zero by construction, it leaves a smaller residual phantom group difference. On real data, it must also leave a group effect that holds up across radii and atlases.
The second competitor is mask erosion: discarding voxels with $e_i$ above a threshold before computing statistics. We run it at matched thresholds and report both the residual phantom effect and how much it biases selection. Atrophy changes which voxels survive erosion, so erosion can manufacture its own group difference.
For downstream detection, the strongest existing method is the deep dynamic-connectivity attention model of Wei et al.. It reports 66.10% sensitivity and 76.76% accuracy for asymptomatic AD on EHBS, where an SVM reached 0% sensitivity, and 93.76% accuracy on ADNI. Ada-FCN reports 79.68% ADNI accuracy, but for a different task definition, so we do not treat it as directly comparable. We will retrain Wei et al.'s architecture on our ADNI split with raw versus boundary-corrected input maps, using the same hyperparameter search budget for both.
Success is not measured by beating the reported accuracy. It means that corrected inputs keep most of the detection sensitivity while geometry-only surrogate inputs lose it. That pattern would show the model is reading function, not shape.
The key ablation swaps the geometry-matched null for a null simulated on a template mask shared by all subjects. This isolates whether subject-specific geometry is the active ingredient.
Why not just…?
Why not just regress out cortical thickness? This is the most serious threat. Thickness is a parcel-level average. Boundary exposure depends on folding, sulcal widening and partial volume at the voxel level, and it relates to the coherence statistic nonlinearly, because fewer neighbors raise variance as well as shift the mean. The surrogate experiment settles the question directly. If thickness regression removes the phantom effect as completely as the geometry-matched null, we will say so, and the contribution reduces to showing the artifact exists and that a cheap fix suffices. That is still useful, but it would mean the estimator itself is unnecessary.
Why not just use isotropic smoothing or carefully tuned smoothing? Functional Brain Image Analysis Using Joint Function-Structure Priors shows that Gaussian smoothing bleeds signal across anatomical boundaries and shifts peaks. Cross-validated bandwidth selection (Optimal spatial regularisation of autocorrelation estimates in fMRI analysis) tunes a global kernel and regularizes temporal autocorrelation, not neighborhood truncation. Smoothing mixes fluid and white-matter signal into edge voxels, and does so more in atrophied brains, so it should worsen the artifact. We include smoothed pipelines in the surrogate test to check this.
Why not just erode masks or work on the cortical surface? Erosion is a fair test and is run as a baseline. Its weakness is selection bias tied to atrophy, which we quantify. Surface-based analysis removes cross-sulcal mixing, but it keeps parcel borders and the medial-wall edge. Our exposure map and null apply on the surface mesh unchanged, so the surface is an additional setting for the method, not a replacement for it.
Decisive Checks
The first check is the phantom effect on null surrogates. If the amyloid-group difference in uncorrected local Moran's I on null surrogates is negligible relative to the observed real-data difference, the core hypothesis is wrong and we stop there.
If the phantom effect is substantial, the second check asks whether the real-data dissociation survives correction. Specifically, does a corrected functional group difference remain while structural measures show no significant group difference? We also test whether boundary exposure itself differs by amyloid status, which would identify the pathway by which atrophy enters the functional measure.
The third check asks whether the DMN-versus-control-network ranking shifts in proportion to each network's mean exposure.
The fourth check, a spatial-dependence sanity test, recomputes group-level inference with Moran's I applied to regional residuals. If those residuals are spatially autocorrelated, the naive tests are anticonservative, and we re-estimate significance with a spatial-lag model. The same treatment is applied to the 21.6% fractional-anisotropy mediation result of Li et al., as a test of whether ignoring dependence among neighboring fiber bundles overstates that estimate.
References
- Results on finite wireless sensor networks: Connectivity and coverage
- Connectivity in secure wireless sensor networks under transmission constraints
- Performance Analysis of Connectivity and Localization in Multi-Hop Underwater Optical Wireless Sensor Networks
- Parametric estimation for Gaussian fields indexed by graphs
- Cube Versus Torus Models and the Euclidean Minimum Spanning Tree Constant
- Spectral analysis of virus spreading in random geometric networks
- Sharp Exact Recovery Threshold for Two-Community Euclidean Random Graphs
- Ada-FCN: Adaptive Frequency-Coupled Network for fMRI-Based Brain Disorder Classification
- mBrainGT: Modular Graph Transformer for Brain Disorder Diagnosis
- A deep spatio-temporal attention model of dynamic functional network connectivity shows sensitivity to Alzheimer's in asymptomatic individuals
- Functional Connectome Signatures of Patients with Asymptomatic and Typical Alzheimer’s
- Mapping Structural Connectome Patterns in Alzheimer's Disease
- Functional Brain Image Analysis Using Joint Function-Structure Priors
- Optimal spatial regularisation of autocorrelation estimates in fMRI analysis