Brain disorders are rare and varied, and their labels are noisy. A useful detector should therefore learn what healthy brains look like and flag departures from that, without needing a labeled example of every disorder. Current brain anomaly detectors learn "normal" entirely from functional signals, meaning the correlations between regional activity time series. That leaves them free to model scanner noise, motion, and site quirks as if they were normal brain behavior. This proposal takes a different starting point. White-matter wiring physically constrains how regions can communicate. So the part of a person's functional activity that does not follow the wiring should be small and stable in healthy people, and it should grow in disease. We propose to measure exactly that part and score anomalies by how unusual it is relative to a population of controls.
Concretely, we treat each fMRI frame as a signal sitting on a structural brain graph and split it into graph frequencies. Low frequencies are activity patterns that vary smoothly along white-matter tracts; high frequencies are patterns that cut against them. We summarize each subject by how their functional energy is spread across these frequencies within each canonical brain network. We then fit a deliberately low-capacity, contamination-robust density model to healthy controls. The bet is that this representation matters more than detector capacity. In the realistic clinical regime, a small robust model on structure-decoupled features should beat deep graph-level detectors and functional-only temporal hypergraph models. That regime has two parts: control pools contain undiagnosed cases, and the anomalies of interest are subtle ("near" the normal distribution).
Core Claim
The claim is that structural connectivity, used as the basis in which functional activity is measured rather than as an extra input channel, gives a representation where subject-level connectome anomalies are easier to separate and harder to contaminate. The claim is falsifiable in one clean way. Suppose the same pipeline performs just as well with a degree-preserving randomly rewired structural graph, or with no structural graph at all. Then the structural prior is doing nothing, and the idea fails.
This combines two lines of work that have not met. Graph signal processing work shows that structure-constrained functional components carry biologically meaningful variance. Functional brain activity constrained by structural connectivity reveals cohort-specific features for serum neurofilament light chain decomposes fMRI on a 66-region tractography graph and explains 42.26% of serum neurofilament light chain variance in controls with a supervised linear regression. Brain Functional Connectivity Estimation Utilizing Diffusion Kernels on a Structural Connectivity Graph uses heat diffusion on diffusion-MRI graphs to suppress spurious functional edges when recovering subnetworks in healthy people. Both motivate the representation, but neither detects anomalies or trains only on normal data.
On the anomaly side, Learning Temporal Higher-order Patterns to Detect Anomalous Brain Activity (HADiB) learns temporal hypergraphs purely from functional time series with contrastive corruption. Anomaly Detection in Multiplex Dynamic Networks: from Blockchain Security to Brain Disease Prediction (ANOMULY) detects edge anomalies with a generic edge-swapping objective, tested on 40 ADHD subjects. Neither uses anatomy. The obvious combination would bolt diffusion-MRI edges onto HADiB as an extra layer of messages between regions. We do not do that. We change what is scored, and we keep the detector simple enough that any gain can be attributed to the representation.
Research Program
Let $A$ be a structural connectivity matrix on $N$ regions, $L = D - A$ its graph Laplacian with degree matrix $D$, and $U$ the Laplacian eigenvectors ordered by eigenvalue. For a regional fMRI frame $x_t \in \mathbb{R}^N$, the graph Fourier coefficients are $\hat{x}_t = U^\top x_t$. We partition the eigenvalues into $B$ frequency bands and, for each canonical network $k$ (default mode, somatomotor, dorsal attention, visual, and the rest of the atlas), compute the time-averaged share of energy in band $b$:
$$ e_{k,b} = \frac{1}{T}\sum_{t=1}^{T} \frac{\lVert P_k U_b \hat{x}_{t,b} \rVert^2}{\lVert P_k x_t \rVert^2} $$
Here $U_b$ and $\hat{x}_{t,b}$ are the eigenvectors and coefficients in band $b$, $P_k$ restricts to regions in network $k$, and $T$ is the number of frames. The subject descriptor $e \in \mathbb{R}^{KB}$ is a few dozen numbers long. The ratio form cancels global signal amplitude, which is one of the main site-dependent quantities.
The anomaly score is a robust Mahalanobis distance to the control population:
$$ s(e) = (e - \mu_h)^\top \Sigma_h^{-1} (e - \mu_h) $$
Here $\mu_h$ and $\Sigma_h$ are the mean and covariance fitted to the control pool with iterative trimming. We fit, drop the controls with the largest scores up to an assumed contamination fraction, and refit. This is what defends against undiagnosed patients in the "healthy" pool.
Because $e$ decomposes by network and band, the per-network terms of $s$ localize the anomaly without extra machinery. This matters for the clinical question. We expect the pattern to be decoupling concentrated in the four networks where Disrupted intrinsic functional brain topology in patients with major depressive disorder found nodal efficiency drops. That study found reduced global efficiency ($t=-2.601$) and local efficiency ($t=-2.771$) in major depressive disorder across 16 sites.
The main practical obstacle is that most clinical fMRI cohorts, including REST-meta-MDD, lack per-subject diffusion MRI. We therefore build a template structural connectome in the Dosenbach 160-region and AAL parcellations from Human Connectome Project diffusion data, and use it as the shared basis. Where per-subject diffusion exists (HCP, and the concussion cohort of the neurofilament paper), we also measure whether individual wiring adds anything beyond the template. A secondary, learned variant replaces the Mahalanobis step with a small normalizing flow on $e$. This tests whether any detector capacity helps once the representation is fixed.
We will also report a failure mode we expect. Head motion adds spatially rough signal, which will load on high graph frequencies and could masquerade as decoupling. To guard against this, every experiment reports the correlation between $s$ and framewise displacement, and repeats the main comparison with motion-matched subsets.
Baselines
The naive reference point is the robust Mahalanobis score applied to classical whole-brain topology summaries: global efficiency, local efficiency, modularity, and per-network nodal degree. These are computed from static Pearson correlation matrices, with the same trimming and control pools as our method. If our score cannot beat this, the graph-frequency representation adds nothing over textbook graph theory, which already separates recurrent depression from controls at the group level.
The second baseline is the same robust Gaussian and a small normalizing flow applied to the flattened lower-triangular correlation matrix, after PCA to the same dimensionality as $e$. This is the unsupervised analogue of BrainNetMLP: An Efficient and Effective Baseline for Functional Brain Network Classification. That paper shows a plain MLP on flattened correlation matrices matches or beats graph transformers in supervised classification (+2.1% accuracy and +1.2% AUROC on HCP over the second-best model). Of all the competitors, this "flattened correlation plus simple density model" is the one most likely to match us.
The strongest deep brain-specific competitor is HADiB. The file reports it at 92.23% AUC on ADHD hyperedge anomaly detection, against 86.23% for the best prior baseline, ADMire (Learning Temporal Higher-order Patterns to Detect Anomalous Brain Activity). The file also cites a figure of about 79% AUC on Parkinson's ROI-level detection, but attributes it inconsistently, once to HADiB and once to USAD, so we do not rely on it. HADiB's reported task is hyperedge-level rather than subject-level. We will therefore run it from its released design on our splits, with subject scores formed by averaging hyperedge scores, and give it the same number of tuning trials selected on held-out controls only.
The strongest general graph-level detectors are FANFOLD (FANFOLD: Graph Normalizing Flows-driven Asymmetric Network for Unsupervised Graph-Level Anomaly Detection) and SIGNET. FANFOLD ranked first on 9 of 15 TUDataset graph sets. SIGNET had the best average AUROC rank, 3.37, across 35 datasets in Unifying Unsupervised Graph-Level Anomaly Detection and Out-of-Distribution Detection: A Benchmark. We run both on thresholded correlation graphs with regional time-series statistics as node features. No reported result was found on Lacuna for any of these methods on REST-meta-MDD, ABIDE or ADHD subject-level anomaly detection, so all numbers there will be produced by us under matched budgets.
The key ablation keeps everything fixed but swaps the basis $U$. The alternatives are a degree-preserving rewired template, the eigenvectors of the subject's own functional correlation matrix (a function-only spectral basis), and the identity matrix (no graph). Beating the baselines means a higher leave-site-out AUROC on recurrent depression versus controls, with a bootstrap confidence interval excluding zero. It also means a smaller AUROC drop than every deep baseline as contamination rises.
Why not just…?
Why not just add diffusion-MRI edges as an extra layer in HADiB or ANOMULY? That is the obvious combination, and it may help. But it confounds two things: more input information and more parameters trained against heuristic corruptions. Heuristic corruptions are exactly what UB-GOLD finds brittle on near-distribution shifts. We run this variant, HADiB with template-structure messages, as an additional baseline. If it matches our method under contamination, the representation claim weakens to "structure helps any model," which is still worth knowing but is a smaller result.
Why not just use classical graph metrics with a one-class SVM or Isolation Forest? UB-GOLD reports that two-step embedding-plus-classical-detector pipelines lose to end-to-end deep models on generic graphs. However, those were high-dimensional learned embeddings, not a few dozen neurobiologically chosen features. Our naive baseline is exactly this approach, and it is a real threat on recurrent depression, where group differences are already significant. Our specific prediction is that it fails on first-episode drug-naive patients, whose topology the depression study found preserved. If structure-decoupling also fails there, we will say so.
Why not just fit a density model to the flattened correlation matrix? This is the most serious alternative, because BrainNetMLP-style results show flattened correlations carry most of the discriminative signal. Its likely weakness is contamination and site variability. About 12,700 correlation entries estimated from roughly 1,500 controls give a covariance dominated by noise, and a few contaminating patients shift it easily. The contamination curve and the leave-site-out split decide this directly. If the flattened model ties us on both, we will conclude that the gain comes from dimensionality reduction rather than anatomy, and the rewired-basis ablation will show the same thing.
Why not reconstruction error from a graph autoencoder? Brain subtle anomaly detection based on auto-encoders latent space analysis : application to de novo parkinson patients found that reconstruction error misses subtle early-Parkinson's changes, while latent-density scoring reached a median g-mean of about 0.65. Anomaly Transformer: Time Series Anomaly Detection with Association Discrepancy reports an 18.76% F1 gain from abandoning reconstruction scoring. Both point toward density scoring on a well-chosen representation, which is what we do. We still include a graph autoencoder with residual scoring for completeness.
Proposed Experiments
The primary testbed is REST-meta-MDD: 821 patients and 765 controls from 16 sites, 160 Dosenbach regions. We use leave-one-site-out evaluation, with controls from the held-out site split so that half calibrate site-wise standardization and half are tested. Recurrent depression is the "far" anomaly and first-episode drug-naive depression the "near" one. This gives a clinically grounded near-versus-far test that UB-GOLD lacks, since it contains no connectome data.
We then transfer the same pipeline without retuning to ABIDE and ADHD. These are cohorts where HADiB was evaluated, so the transfer tests whether the score generalizes across disorders. HCP supplies the template structural graph and a large healthy calibration pool. The concussion cohort, with per-subject diffusion data, tests individual versus template structure.
Contamination is injected by moving held-out patients into the training pool at 0, 5, 10 and 20%. We report AUROC, AUPRC, and the slope of AUROC against contamination for every method. Localization is checked by whether per-network score terms concentrate in the default mode, somatomotor, dorsal attention and visual networks, compared against the nodal findings of the depression study.
The decisive checks are three.
First, the real template basis must beat the rewired and function-only bases on first-episode depression AUROC. If it does not, the anatomical claim is false.
Second, our contamination slope must be flatter than that of FANFOLD, SIGNET and HADiB at matched clean-data AUROC. If it is not, the robustness claim is false.
Third, the score must stay predictive after regressing out framewise displacement. If it does not, we are detecting motion rather than disease.
References
- Functional brain activity constrained by structural connectivity reveals cohort-specific features for serum neurofilament light chain
- Brain Functional Connectivity Estimation Utilizing Diffusion Kernels on a Structural Connectivity Graph
- Learning Temporal Higher-order Patterns to Detect Anomalous Brain Activity
- Anomaly Detection in Multiplex Dynamic Networks: from Blockchain Security to Brain Disease Prediction
- Disrupted intrinsic functional brain topology in patients with major depressive disorder
- BrainNetMLP: An Efficient and Effective Baseline for Functional Brain Network Classification
- FANFOLD: Graph Normalizing Flows-driven Asymmetric Network for Unsupervised Graph-Level Anomaly Detection
- Unifying Unsupervised Graph-Level Anomaly Detection and Out-of-Distribution Detection: A Benchmark
- Brain subtle anomaly detection based on auto-encoders latent space analysis : application to de novo parkinson patients
- Anomaly Transformer: Time Series Anomaly Detection with Association Discrepancy