When small blood vessels in the brain slowly fail, patches of damaged white matter appear on MRI. These patches, called white matter hyperintensities, are the visible trace of chronic cerebral hypoperfusion, a long-term shortfall in blood supply to the brain. Models that predict cognition from these maps usually treat each damaged voxel as an independent input with a fixed cost. Physiology suggests a different picture. White matter tracts draw blood from a small number of shared arterial supply routes. Animal work shows the brain tolerates blood-flow loss up to a threshold and then fails sharply. If that is right, damage to a given tract should cost more when the other tracts on the same supply route are also damaged, because the shared supply is already near its limit.
Computer systems have a close analogue, called the "noisy neighbor" effect. Programs sharing a cache, memory bus, or power budget run normally until the shared resource saturates, and then slow down sharply. Systems researchers have learned to predict this slowdown with very small parametric models. We propose to borrow that modeling style for lesion-symptom mapping. The model asks a specific question: is the cognitive cost of damage in a tract amplified, in a saturating way, by the total damage load in the same vascular territory? The model is designed to separate two things. One is the shared, territory-level damage that makes voxel features highly redundant. The other is the tract-specific damage that carries strategic signal.
Core Claim
Voxel-wise lesion maps are redundant mainly because voxels in the same supply territory are damaged together. That redundancy is not just noise to be compressed away. It measures how heavily a shared resource is loaded, and that load should gate the effect of tract-specific damage. A model with a few parameters per territory and an explicit saturating interaction will match or exceed flexible voxel-wise models on held-out cognitive prediction. It will also give a falsifiable answer to whether the interaction exists. The decisive test is anatomical. The gain over an additive model should appear when lesion load is grouped by arterial territories, and it should vanish when voxels are grouped into random parcels of matched size.
Grounding in Prior Work
Three papers motivate the proposal by showing that spatial lesion structure carries signal that simple summaries miss. Individualized lesion-symptom mapping using explainable artificial intelligence trains a convolutional network on 3D lesion masks from the TRACE-VCI cohort and explains its predictions by occluding parts of the input. On real attention and executive scores, total lesion volume reaches only $R^2 = 0.013$, while support vector regression reaches $0.291$ and the network $0.216$. That paper treats voxels as independent pixels and probes them by local occlusion, which cannot reveal an interaction between distant regions on the same supply route. Our model tests exactly that interaction.
Signature White Matter Hyperintensity Locations Associated With Vascular Risk Factors analyses 15,653 individuals with linear mixed models. It shows that hypertension leaves an anterior-predominant regional signature that survives correction for total volume, whereas most other risk factors only raise overall load. We adopt its key move, separating regional signal from global collinearity. We replace its linear correction with a territory-level gate that can express thresholds. Cortical microinfarcts in memory clinic patients are associated with reduced cerebral perfusion finds that small cortical infarcts accompany blood-flow reductions across the whole anterior circulation, not just around each lesion. This supports treating the territory, rather than the lesion, as the unit of shared vulnerability.
Two papers constrain the design. Deeper cerebral hypoperfusion leads to spatial cognitive impairment in mice reports no memory deficit below roughly a 30% blood-flow reduction and deficits beyond it. That argues for a saturating rather than linear gate. Lesion load is only a proxy for blood flow, however, so the threshold must be estimated from data rather than fixed at 30%. Preserved Transcallosal Inhibition in Nondemented Elderly Patients with Leukoaraiosis finds that early executive dysfunction tracks frontal injury rather than global disconnection. A global-load model would therefore be too coarse, which motivates gating at the territory level.
The method comes from systems work on shared-resource contention. iGniter predicts how much slower GPU inference runs when other programs share the chip. It needs only 11 profiling measurements per network, with 1.53% to 5.02% error, by modeling slowdown as a program's solo cost amplified by a few named sources of interference. VELTAIR shows that a single interference signal, the rate at which a program's data is evicted from the shared cache, is enough to steer scheduling. We take from these papers the shape of the model: a small amplification term driven by the load on each shared resource. We take nothing about hardware. Unlike variance-based compression such as spatial independent component analysis, used in Large-scale functional network connectivity mediate the associations of white matter lesions, our shared factor is defined by anatomy and supervised by outcome. Gaussian Image Anomaly Detection with Greedy Eigencomponent Selection shows that the highest-variance directions in feature spaces often carry background rather than task signal. In lesion maps, the top component is very likely just total volume, which we already know is nearly uninformative.
The Model
For patient $i$, let $x_{ir}$ be the fraction of white matter tract $r$ that is lesioned. Let $L_{it}$ be the lesion load in arterial territory $t$, and let $T(r)$ be the territory supplying tract $r$. The prediction for cognitive score $y_i$ is
$$\hat y_i = \beta_0 + \gamma L_i^{\text{tot}} + \sum_r w_r\, x_{ir}\left[1 + \alpha_{T(r)}\,\sigma\!\left(\frac{L_{i,T(r)} - \tau_{T(r)}}{s}\right)\right],$$
where $L_i^{\text{tot}}$ is total lesion volume and $w_r$ is the base cost of tract $r$. For each territory $t$, $\alpha_t \ge 0$ is the amplification once that territory saturates, and $\tau_t$ is its threshold load. The function $\sigma$ is the logistic function, with shared softness $s$. Setting every $\alpha_t = 0$ gives the additive ablation.
The territory map is the main assumption. We use a standard arterial territory atlas, with watershed zones treated as their own territories. Tract weights are sparsity-regularized. The model has on the order of 50 tract weights and three parameters per territory, far fewer than a voxel-wise network. This matters for cohorts of a few hundred patients.
Research Program
The work has three stages.
First, we extend the simulation protocol of the lesion-symptom mapping paper. That paper generated synthetic scores on 821 real lesion maps, driven by known regions. We add a second family of synthetic scores in which the cost of strategic regions is amplified by load in their territory. We then measure two things: whether each method recovers the true interaction, and whether occlusion-based attribution maps misplace it.
Second, we fit all models to TRACE-VCI attention and executive scores under identical nested cross-validation.
Third, we test external meaning. In the Meta VCI Map data, the anterior signature of hypertension predicts that anterior territories should have lower estimated thresholds $\tau_t$ in hypertensive than in normotensive participants. In the asymptomatic small vessel disease cohort, the model predicts that territory saturation should explain the reported between-network connectivity changes beyond what total lesion volume explains. Those connectivity changes are between the executive control and salience networks, and between the auditory and posterior default mode networks. In Medea-7T, cortical microinfarct presence serves as an independent marker of territory hypoperfusion. If the gate tracks real supply stress, patients with microinfarcts in a territory should sit further above that territory's threshold.
Access to TRACE-VCI and Meta VCI Map requires consortium collaboration. If only RUN DMC lesion maps are available, the simulation stage still runs in full.
Baselines
The naive reference is linear regression on total lesion volume. Its reported $R^2$ on TRACE-VCI real scores is $0.013$ (lesion-symptom mapping paper), and any proposed model must clear it by a wide margin.
The strongest existing method is voxel-wise support vector regression on lesion masks. The same paper reports $R^2 = 0.291$ on TRACE-VCI attention and executive scores, ahead of its own convolutional network at $0.216$. We will rerun this baseline ourselves rather than compare to the published figure. It will use the same folds, the same lesion preprocessing, and a nested search over kernel and regularization with the same number of configurations given to our model. We will also rerun the convolutional network under that protocol as the flexible deep competitor. Beating the support vector baseline means a higher pooled held-out $R^2$ with a paired bootstrap confidence interval on the difference that excludes zero. We pre-specify a gain of at least 0.03 as practically meaningful. Matching its accuracy with far fewer parameters and an interpretable interaction would be a weaker but still useful result.
The linear residualization approach of the Meta VCI Map paper is reimplemented as tract volumes plus a total-volume covariate. It represents the best purely additive regional model.
Gradient-boosted trees on territory and tract volumes are the flexible interaction learner. We found no reported result on Lacuna for this baseline on these cohorts.
Two ablations isolate the key ingredient. The first sets all $\alpha_t = 0$. The second keeps the gate but replaces arterial territories with random contiguous parcels of matched size, repeated many times to form a null distribution.
Why not just…?
Why not just let a voxel-wise model learn any interaction on its own? A convolutional network can in principle represent territory-gated effects. On real TRACE-VCI data, though, it already underperforms support vector regression (0.216 vs 0.291), which suggests 813 patients are too few to discover such structure without guidance. The simulation stage settles this directly. If the network recovers planted interactions as well as our model at TRACE-VCI sample sizes, our structural prior adds nothing, and we would report that.
Why not just use gradient boosting on territory volumes? This is the most serious threat. Boosted trees on the same anatomical inputs can learn thresholds and interactions without our specific functional form. If they match our accuracy, the anatomical grouping rather than the contention-style gate is doing the work. The study would still deliver the territory finding, but the borrowed model form would not be justified. We will report this outcome plainly. Our model's remaining advantage would then be its direct, testable threshold parameters, which the external Meta VCI Map and Medea-7T checks either support or fail to support.
Why not just compress lesion maps with PCA or ICA first? Variance-based components in lesion maps are dominated by overall load. Total load alone gives $R^2 = 0.013$, and the greedy eigencomponent work shows high-variance directions often carry background rather than task signal. We include principal-component regression with the component count chosen by nested cross-validation. We expect it to trail, and the gap measures how much an anatomy-defined shared factor beats a variance-defined one.
Decisive Checks
Three results would establish the idea. First, the territory-gated model must beat the additive ablation on held-out TRACE-VCI scores. Second, that advantage must disappear under random parcellations, with the arterial atlas falling in the extreme tail of the null distribution. Third, the estimated thresholds must line up with independent markers of hypoperfusion, namely hypertension status in Meta VCI Map and microinfarct territory in Medea-7T.
If the gain persists under random parcels, the effect is generic lesion co-occurrence, not shared supply, and the vascular interpretation fails. If the gated model beats the additive one but not gradient boosting, the anatomy matters but the contention form does not. Either outcome is informative, which is why these checks come before any claim about clinical use.
References
- Individualized lesion-symptom mapping using explainable artificial intelligence for the cognitive impact of white matter hyperintensities
- Signature White Matter Hyperintensity Locations Associated With Vascular Risk Factors Derived From 15 653 Individuals
- Cortical microinfarcts in memory clinic patients are associated with reduced cerebral perfusion
- Deeper cerebral hypoperfusion leads to spatial cognitive impairment in mice
- Preserved Transcallosal Inhibition to Transcranial Magnetic Stimulation in Nondemented Elderly Patients with Leukoaraiosis
- iGniter: Interference-Aware GPU Resource Provisioning for Predictable DNN Inference in the Cloud
- VELTAIR: towards high-performance multi-tenant deep learning services via adaptive compilation and scheduling
- Large-scale functional network connectivity mediate the associations of white matter lesions with executive functions and information processing speed in asymptomatic cerebral small vessels diseases
- Gaussian Image Anomaly Detection with Greedy Eigencomponent Selection