Disconnection Coherence: A Rater-Robust Spatial Statistic for Forecasting Disability in Multiple Sclerosis

In multiple sclerosis (MS), how much visible lesion a scan shows says surprisingly little about how disabled the patient is, or how fast they will get worse. This mismatch is called the clinico-radiological paradox. The strongest recent attempt to forecast worsening fed atlas-based regional brain volumes into five standard classifiers. Those models detected current disability well, but no model predicted two- or five-year worsening better than chance. Separately, group-level studies show that damage spreads along wiring: a white-matter lesion is followed by shrinkage in the gray-matter nuclei that its fiber tract connects to. Volume summaries throw away exactly this "lesion upstream, damage downstream" pattern. They are also hostage to radiologists, because two experts outlining the same lesions disagree substantially.

We propose a single patient-level number that measures whether a patient's damage follows the brain's wiring. The tool is spatial autocorrelation, the statistics epidemiologists use to ask whether disease cases cluster. Here it is computed on a graph whose nodes are white-matter tracts and deep gray-matter nuclei and whose edges are anatomical tract-to-nucleus connections. A high value means that the tracts carrying the most lesion are the ones whose connected targets show the most tissue damage. We call this disconnection coherence. It is a ratio of covariances, so a rater who systematically draws every lesion larger, or smaller, leaves it unchanged. The idea in two sentences: patients whose damage is already propagating along connections should progress faster than patients with the same lesion load scattered incoherently. A wiring-aware correlation statistic can capture this in one interpretable, rater-robust number, where volume snapshots cannot.

Core Claim

The claim has three parts, and each is testable on its own. First, a tract-graph version of Moran's $I$ separates network-coherent from incoherent damage within individual patients. Moran's $I$ is the standard index of spatial clustering. The bivariate version below asks whether high lesion load on one node co-occurs with high damage on the nodes wired to it:

$$I_{\ell a} = \frac{N}{\sum_{i,j} w_{ij}} \cdot \frac{\sum_{i,j} w_{ij}\,(\ell_i-\bar\ell)(a_j-\bar a)}{\sqrt{\sum_i (\ell_i-\bar\ell)^2}\,\sqrt{\sum_j (a_j-\bar a)^2}}$$

In this formula, $\ell_i$ is the fraction of tract $i$ occupied by lesion. $a_j$ is the damage z-score of node $j$ relative to healthy-control norms, measured as volume loss in a nucleus or quantitative T1 change in normal-appearing tissue. $w_{ij}$ is the normative anatomical connection weight between nodes $i$ and $j$, and $N$ is the number of nodes. To get a patient-level coherence z-score, we compare the observed $I_{\ell a}$ against a null distribution built by shuffling node values within each compartment, so tracts are shuffled among tracts and nuclei among nuclei.

Second, the statistic is unchanged when lesion loads are rescaled by any positive factor and shifted by a constant. Uniform over- or under-segmentation by one rater therefore cancels exactly. Only spatially uneven disagreement can move it. We expect coherence to show markedly higher between-rater agreement than lesion volume, which carries every rater's bias in full.

Third, coherence adds prognostic information about future worsening beyond lesion volume and beyond tract-wise mean damage. The latter is the strongest existing tract-level predictor. The model that makes the forecast is a sparse additive model that admits nonlinearity only when it earns it. Its explanation is a few named spatial statistics plus a local map of which tract-to-nucleus edges drive the score, rather than a voxel saliency map.

Relation to Prior Work

Two papers motivate the gap. Evaluation of machine learning-based classification of clinical impairment and prediction of clinical worsening in multiple sclerosis trained five classifiers on Brainnetome parcellations, with SHAP rankings for explanation, in 455 patients. They reached an AUC of 0.83 ± 0.07 for current disability (EDSS ≥ 4) and found no significant prediction of worsening ($p > 0.05$). We reuse their setup as a baseline. We differ by adding a statistic of spatial coupling rather than more regional volumes. White-matter lesions drive deep gray-matter atrophy in early multiple sclerosis showed with voxel-wise group regression in 249 patients that lesions in specific tracts predict atrophy in connected deep gray matter. We turn that group-level finding into a per-patient score and ask whether it predicts the future.

Two papers supply methods. Tract-wise microstructural analysis informs on current and future disability in early multiple sclerosis built normative quantitative-T1 z-score maps from 92 controls and showed that tract-averaged damage in normal-appearing white matter, especially in brainstem and cerebellar tracts, tracks two-year EDSS change. We take their normative z-scores as our node damage values $a_j$ and their tract means as our hardest baseline. Our contribution must exceed theirs, because tract means already capture where damage is. A reluctant additive model framework for interpretable nonlinear individualized treatment rules fits a sparse linear model first and admits spline terms only when a concordance criterion shows real gain. We adapt that rule from treatment selection to risk prediction.

The inter-observer literature constrains the design. Longitudinal multiple sclerosis lesion segmentation data resource provides 82 scans from 19 patients, each segmented by two raters, and documents substantial boundary disagreement. It reports no clinical link and no algorithm numbers. Multi-rater segmentation models such as OMBA-UNet and the multi-decoder U-Net model disagreement as a distribution but stop at segmentation quality.

Calibrate the inter-observer segmentation uncertainty via diagnosis-first principle goes further: it learns rater weights to maximize downstream diagnosis and gains up to 2.73% AUC on 2D fundus images. We treat it as a competitor, not a component. Where the image-to-outcome link is weak, which is the defining property of the paradox, learning label weights against a small outcome set risks fitting noise. A fixed statistic that is invariant by construction to the dominant form of rater bias is the safer route.

Finally, the spatial-statistics literature warns that the choice of weight matrix changes Moran's $I$, and that brain proximity can be Euclidean or wiring-based. We make that choice an explicit ablation rather than a hidden setting.

Research Program

We build a node graph from a normative tract atlas, the DSI Studio tract atlas used by Ravano et al. The roughly forty tracts are nodes, alongside deep gray-matter nuclei, with connection weights taken from where each tract terminates. Each patient's lesion masks are projected onto tracts to give $\ell_i$. Normative z-scores give $a_j$ for nuclei, which are atrophy or T1 measures, and for normal-appearing tract tissue.

From these we compute three statistics. The first is univariate Moran's $I$ of lesion load over the graph, which measures whether lesions cluster in wired-together tracts. The second is the bivariate coherence $I_{\ell a}$ above. The third is its local decomposition, which assigns each edge a contribution to the score and serves as the patient-specific explanation.

The prognostic model is a reluctant additive survival or classification model. It takes a small, pre-registered feature set: baseline clinical variables, lesion volume, thalamic volume, tract-mean damage, and the coherence statistics. To keep the comparison with Noteboom et al. clean, we use their two outcome definitions and cross-validation scheme. Because fast progressors are rare, the primary metric is the concordance index for time to confirmed worsening, with bootstrap intervals. We pre-register the comparisons so that a small cohort cannot be mined for a positive result.

Data come in two tiers, and we state access plainly. Rater robustness is tested on the public two-rater longitudinal MS resource and, as a cross-check of the invariance argument outside MS, on the multi-rater brain-tumor tasks of QUBIQ. The prognostic claim needs an early-MS cohort with baseline structural and ideally quantitative-T1 MRI plus two- to five-year EDSS follow-up. The cohorts described by Noteboom, Ravano and Mühlau are of this kind but are not public, so this tier requires a data-sharing collaboration. The prognostic test is the proposal's main dependency.

Baselines

The naive reference point is a clinical-only model using baseline EDSS, age, disease duration and sex, plus total lesion volume. This is what a neurologist already has. A second reference is the Noteboom et al. pipeline rerun on our cohort: regional volumes with SVM, random forest and XGBoost, explained by SHAP. For current disability it reports an AUC of 0.83 ± 0.07. For worsening it reports no significant prediction. No numeric worsening AUC is given in the file, so beating it means reaching a confidence interval on the concordance index that excludes 0.5 where it does not.

The strongest competitor is tract-wise mean damage in the style of Ravano et al.: normative T1 z-scores averaged per tract, together with tract-weighted lesion load, fed to the same reluctant additive model. It is strongest because it already encodes where damage sits, which is the information our statistic claims to exploit. The file reports that it significantly improved correlation with ΔEDSS but gives no correlation magnitude, so no reported result was found on Lacuna for its effect size. Every model, including ours, gets the same folds, the same nested tuning budget, and the same feature-selection rule. Beating this baseline means a pre-registered improvement in the concordance index whose paired bootstrap interval excludes zero.

The ablation that isolates the key ingredient keeps every feature fixed and changes only the weight matrix $w_{ij}$. It runs three versions: the anatomical tract-to-nucleus graph, an inverse-Euclidean-distance graph over the same nodes, and degree-preserving random rewirings of the anatomical graph. If the Euclidean or rewired graphs match the anatomical one, the gain comes from generic spatial smoothness rather than wiring, and the central claim fails.

A fourth baseline is a 3D convolutional network trained end to end on baseline images, explained with Grad-CAM. It represents the "let the network learn space" view.

Why not just…?

Why not just use tract-wise mean damage and tract-weighted lesion load? This is the real threat. An additive model over per-tract lesion loads and per-target damage scores captures each location's main effect. It does not capture the patient-specific coupling between a lesioned tract and its own targets. That coupling would require on the order of hundreds of tract-by-nucleus interaction terms, which a cohort of a few hundred cannot support. Coherence is a pre-specified, one-number compression of those interactions. The decisive comparison is the nested model. If adding coherence to tract means plus tract loads yields no gain in the concordance index, we will report that the spatial-coupling hypothesis adds nothing beyond Ravano et al.

Why not just train a 3D convolutional network and inspect saliency? Convolutions see local grid neighborhoods, while the dependency we target links distant regions through axons. The outcome is also weak and rare, which is the regime where Noteboom et al. found even simple models could not forecast worsening. Saliency methods are known to be noisy and unstable (Opti-CAM; saliency-guided training). Agreement with plausible regions also does not establish faithful reasoning: Haddada et al. found an expert-alignment Kendall correlation of −0.39 for one model. We run the network as a baseline under identical folds. If it wins, that is an honest result, and we would test whether its Grad-CAM maps themselves show tract-graph coherence.

Why not just learn rater weights for the outcome, diagnosis-first style? That approach worked on 2D benchmarks where the segmented structure determines the diagnosis. The paradox means visible lesion boundaries only weakly determine EDSS. Fitting label weights against a few hundred noisy outcomes is likely to overfit. We implement a 3D adaptation, with rater-weighted soft masks feeding the same downstream model, and compare its out-of-fold concordance and its across-fold weight stability with the fixed statistic.

Why not just average multiple raters, or use a multi-rater model's soft mask? Averaging is compatible with our approach, and we will compute coherence on consensus and soft masks too. Averaging does not remove a rater's systematic bias from lesion volume. Our invariance argument says coherence is insensitive to that bias whichever mask is used. The reliability experiment measures directly how much each choice matters.

Proposed Experiments and Decisive Checks

The first check is rater robustness on the two-rater MS resource. For each scan we compute lesion volume, tract-weighted lesion load and coherence from each rater's mask separately, then report between-rater intraclass correlation with bootstrap intervals. We also test invariance directly by dilating and eroding one rater's masks by fixed amounts and tracking how each statistic moves. Lesion volume should shift with every dilation step while coherence stays close to flat. If coherence is no more reliable than volume, the rater-robustness argument fails.

Longitudinal pairs in the same resource provide a second check: coherence should be more stable across a year within patient than between patients, which shows it measures a trait-like pattern. The QUBIQ brain-tumor tasks provide a third check: they test whether invariance holds when raters number more than two.

The decisive prognostic test is the pre-registered nested comparison on the clinical cohort. The model sequence is clinical-only, then plus volumes, then plus tract means and tract loads, then plus coherence. We report the concordance index for time to confirmed worsening and AUC for the binary two-year endpoint. The weight-matrix ablation and a negative control accompany it. In the negative control, coherence is recomputed after randomly reassigning which nucleus each tract connects to, and it should lose its predictive value.

A secondary analysis maps per-patient local coherence. It asks whether the edges that drive high-risk scores concentrate in the brainstem and cerebellar pathways that Ravano et al. tied to future worsening. Because that is an independent prior finding, it serves as an external check on whether the explanations make biological sense.

The proposal succeeds if three things hold together: coherence beats volume on rater agreement, it adds predictive value over tract means, and the anatomical graph beats the Euclidean and rewired graphs. Any one of these failing tells us precisely which part of the idea was wrong.