Pruned Root-Cause Search on the White Matter Skeleton

Brain imaging studies routinely ask where two groups of people differ in their white matter, the bundles of nerve fibers that connect brain regions. The standard tool, tract-based spatial statistics (TBSS), works in three steps. It thins every subject's diffusion MRI map down to a shared "skeleton" running through the centers of the fiber tracts. It then runs one statistical test per skeleton voxel. Finally, it corrects for the tens of thousands of tests. The output is a speckled map of significant voxels. A researcher then has to interpret it by hand against an anatomical atlas and decide whether the disease causes a brain-wide shift, damage to a few specific tracts, or both. We propose to recast this as a root-cause localization problem of the kind solved in software operations. The task becomes finding the smallest, coarsest anatomical unit whose change explains the group difference. We make the search exact by pruning it with provable upper bounds on the test statistic.

The technical core is simple. A region's summed test statistic behaves erratically as the region shrinks, so the usual "if a region fails, its sub-regions fail too" shortcut does not apply. Its best achievable value over all sub-regions, however, has a cheap upper bound that can only decrease as we descend the anatomical hierarchy. That bound lets a branch-and-bound search compute, for every permutation of the group labels, the exact maximum statistic over an exponentially large family of connected skeleton regions. This yields family-wise error control over regions rather than voxels, and a hierarchical report such as "global shift plus an extra focal deficit in the anterior corona radiata." The report replaces a voxel map that someone must decode by hand.

What the three directions contribute

The search machinery comes from the literature on pruning non-monotone objectives. Feature Selection in Kernel Space proves that the subgradient contribution of a high-order feature combination is bounded by that of its constituents, which lets whole branches of a kernel feature space be skipped. Bounded coordinate-descent for biological sequence classification does the same for coordinate-descent steps over all substrings. It uses the fact that a longer string cannot be more frequent than its prefix. Both papers bound a gradient, meaning the change in fit from adding one feature, over a lattice of structures. We reuse that logic directly. The score test for adding a "this region differs" indicator to a group-comparison regression is exactly such a gradient evaluated at zero effect. What we change is the lattice: it becomes connected regions of a 3D skeleton graph instead of parse trees or strings.

The problem formulation comes from hierarchical fault localization. PSqueeze localizes failures in online services to the most parsimonious combination of categorical attributes. Its key observation is a "generalized ripple effect": when a root cause fires, every descendant data slice shows the same proportional change. PSqueeze exploits this with bottom-up clustering followed by a heuristic top-down search. It is restricted to categorical attributes, its authors list continuous numerical attributes as an open problem, and its search carries no guarantee that a pruned branch holds no better explanation. We carry over the ripple idea as a parsimony rule. A skeleton node is reported as the root cause if its children share a common effect; otherwise the search descends. We replace the heuristic search with a bounded one. RCAEval motivates this choice. There, detectors specialized to one kind of signal collapse outside their niche: BARO scores 1.0 on disk faults but 0.63 Average@5 on network delays. That is the fragility we want to avoid by making the search exact rather than tuned. The operator-theory papers grouped under this direction in the file are unrelated to the proposal and are not used.

TBSS supplies the application and its concrete weaknesses. Ball et al. simulated treatment effects in 90 preterm neonates and report two findings. A 5% global increase in fractional anisotropy (FA, a measure of how directional water diffusion is, often read as fiber integrity) reaches significance in over half the skeleton for groups of at least 20. When no effect exists, however, about 10% false positives appear for small regional differences, and variability rises sharply below 20 subjects. The ENIGMA schizophrenia study of 4,322 people found its largest effect in the whole-skeleton average FA (Cohen's d = 0.42), with FA reductions in 20 of 25 regions. The anterior corona radiata and corpus callosum were the most pronounced, and the diffusion pattern pointed to myelin loss: mean and radial diffusivity rose while axial diffusivity stayed flat. The question "is this global or focal, and through what mechanism" had to be answered by post-hoc region-of-interest aggregation. Our method is meant to answer it automatically and with error control.

Core Claim

We claim that exact region-level family-wise error control over all connected skeleton regions within an anatomical hierarchy is computationally feasible. Monotone bounds on the region score test make per-permutation maximization tractable. The same search also yields a hierarchical root-cause report that separates global from focal effects. We expect this to improve localization and false-positive control over voxel-wise TBSS in small cohorts without losing power for diffuse effects.

The statistic works as follows. Let $z_v$ be the voxel-level score statistic at skeleton voxel $v$ for the group contrast. For a candidate region $R$, the region statistic is

$$T(R) = \frac{1}{\sqrt{|R|}} \sum_{v \in R} z_v,$$

the standardized sum. Under independence it is the score test for a region-wide shift; permutation handles the real spatial correlation. $T$ is not monotone: shrinking $R$ can raise or lower it. Now sort the positive $z_v$ in $R$ in decreasing order as $z_{(1)} \ge z_{(2)} \ge \dots$ and define

$$B(R) = \max_{k} \frac{1}{\sqrt{k}} \sum_{i=1}^{k} z_{(i)}.$$

For any fixed size $k$, the best subset is the top $k$ voxels. Hence $B(R) \ge T(R')$ for every sub-region $R' \subseteq R$, and $B$ can only decrease as $R$ shrinks. $B$ costs a sort to compute. It is loose only because it ignores the requirement that $R'$ be connected. When $B(R)$ falls below the best connected region found so far, the entire subtree under $R$ is discarded with no loss.

Multimodal signatures enter as a small fixed set of weightings of the per-metric statistics, for example a "myelin-like" contrast of FA down and radial diffusivity up with axial diffusivity flat. Each weighting defines its own $z_v$ and runs through the same search. The signature is thus a label on the reported root cause, not a separate uncorrected analysis.

Research Program

We will first build the hierarchy. Its levels are the whole skeleton; hemispheres and tract systems; atlas tracts (the 25 ENIGMA regions); along-tract segments; and, below that, connected sub-regions of the skeleton voxel graph grown within each segment. The connected-subgraph level is the one that is exponentially large and needs pruning.

Next comes the search. For each permutation of group labels, a best-first branch-and-bound finds the maximum of $T$ over the whole family, using $B$ to prune. The distribution of these maxima across permutations gives a threshold that controls region-level family-wise error exactly.

On top of the search sits the reporting rule. Starting at the root, a node is declared significant if its $T$ exceeds the threshold. It is then tested for effect homogeneity across its children, which is the ripple-effect check. A homogeneous significant node is reported as a root cause, and its descendants are tested only for residual effects beyond the node-level shift. A heterogeneous node passes its budget down to its significant children.

We will analyze how tight the bound is in practice: the fraction of nodes expanded as a function of effect size, cohort size, and skeleton branching. We will also prove that pruning never changes the permutation maxima, which we will verify empirically against exhaustive search on small skeletons.

Baselines

The naive reference is voxel-wise TBSS with Bonferroni or uncorrected thresholds, followed by atlas lookup. It shows how much any correction or structure buys.

The strongest existing competitor is standard TBSS with permutation testing and Threshold-Free Cluster Enhancement (TFCE, a method that boosts each voxel's statistic by the extent and height of the signal cluster around it), as used by ENIGMA and in Ball et al.'s sensitivity study. It is the community default, it is well calibrated at the voxel level, and it already pools spatial neighbors. We will run it on the identical registered skeletons and permutation count, with default TFCE parameters plus a small sweep given the same tuning budget as our method's one parameter, the minimum region size. Reported numbers for this baseline are limited to Ball et al.: a 5% global FA effect is significant in over 50% of the skeleton at group size 20 or more, detection of a 10% change with 10 per group, and a roughly 10% false-positive rate for small regional differences under the null. No reported TFCE runtime or region-level localization accuracy was found on Lacuna. We beat it if, at matched family-wise error, we show lower false-positive region reports under the null at n below 20 and higher localization accuracy for planted focal effects. We must also stay within a few points of its sensitivity to global effects.

The second strong competitor is a region-of-interest analysis on the 25 atlas tracts with Bonferroni correction. This is cheap, interpretable, and close to how ENIGMA reached its conclusions. We will also include along-tract functional modeling in the style of Multivariate Varying Coefficient Models for DTI Tract Statistics, which models diffusion metrics as smooth functions along a tract and tests for group effects. That line of work handles 1D tract profiles but not branching sheet-like skeleton regions or a hierarchy.

Two ablations isolate our ingredients. The first uses the same hierarchy and statistic but restricts the family to contiguous along-tract intervals, which can be enumerated exhaustively without bounds. It tests whether the exponential connected-region family matters at all. The second keeps our search but removes the ripple rule and reports all significant nodes. It tests whether the root-cause parsimony rule, rather than the statistic, is what makes the reports interpretable.

Why not just…?

Why not just run TBSS with TFCE and read the clusters against an atlas? This remains the right default for voxel-level questions, and our method must not lose to it on diffuse effects. The evidence against it for our question is Ball et al.'s roughly 10% false-positive rate for small regional differences in small cohorts. Moreover, TFCE maps do not say whether a pattern like ENIGMA's (global d = 0.42 plus stronger frontal and callosal deficits) is one global shift or several focal ones. The planted-effect simulations below measure exactly this distinction, which the atlas reading does by eye.

Why not just test the 25 atlas regions with Bonferroni? This is a real threat for large, region-aligned effects. It will likely match us when the true effect coincides with an atlas label. It fails by construction when an effect occupies part of a tract or straddles two labels, where averaging dilutes it. The partial-tract planted effects will show whether that failure is large enough to matter. If it is not, the honest conclusion is that atlas regions suffice, and our contribution shrinks to the hierarchical global-versus-focal report.

Why not just enumerate along-tract intervals exhaustively and skip the bounds? If the relevant family is contiguous segments along each tract, the number of candidates grows only quadratically with tract length, and pruning buys speed but no power. We take this seriously; it is the first ablation. The pruning is justified only if connected regions in the branching, sheet-like parts of the skeleton recover planted effects that intervals miss. The exact-versus-interval comparison on non-tubular planted shapes is designed to settle this. If intervals match us, we will report the method as an interval scan with a ripple rule and drop the branch-and-bound claim.

Why not just use a heuristic top-down search like PSqueeze on the atlas tree? A heuristic search can miss the true maximum in some permutations. That silently biases the permutation threshold, so its family-wise error guarantee no longer holds. The RCAEval results, where tuned heuristics swing from 1.0 to 0.63 across fault types, are the software-side warning. We will measure this directly by running a greedy variant of our search and comparing its null error rate with the exact one.

Proposed Experiments

The main testbed uses Human Connectome Project diffusion data, as in Kochunov et al., processed through the ENIGMA-DTI TBSS protocol. We will split subjects at random into pseudo-groups of 10, 20, and 45 per group, following Ball et al.'s design. Under no planted effect, we measure the rate of any false region report for each method; this is the family-wise error. We then plant four kinds of known root causes, each at 2%, 5%, and 10% FA changes: a global shift across the whole skeleton; a focal deficit confined to part of one tract; a deficit shaped as a sheet in a branching region where tubular intervals fit poorly; and a global shift plus a focal deficit. For signature detection, the plants alter FA and radial diffusivity jointly while holding axial diffusivity fixed.

Metrics are family-wise error, power, and region overlap (Dice score) between reported and planted regions. We also score hierarchy correctness, meaning whether the method reports "global," "focal," or "global plus focal" as planted. Efficiency is measured as runtime and nodes expanded per permutation.

The decisive checks are threefold. First, pruned and exhaustive search must give identical permutation maxima on downsampled skeletons. Second, the connected-region family must beat the interval ablation on the sheet-shaped plants. If it does not, the pruning claim is withdrawn. Third, our method must beat TFCE plus atlas reading on hierarchy correctness at n = 10 to 20 while staying within a few points of its sensitivity to global effects.

As a real-data check, if ENIGMA-style schizophrenia data can be obtained through the consortium, we will test whether the method's unprompted report matches the published picture. That picture is a global FA reduction with a myelin-like signature, plus residual focal deficits in the anterior corona radiata and corpus callosum. Recovering it would be the clearest evidence that the root-cause report does automatically what the field currently does by hand.