Worst-Case Brain Connectivity from Diffusion MRI via Markov Decision Processes

Diffusion MRI tracks how water moves in the brain, and tractography turns those measurements into estimated white-matter pathways between regions. Probabilistic tractography repeatedly sends virtual particles ("streamlines") through local fiber-direction estimates and counts where they arrive. Its main failure is not noise but false confidence. On the ISMRM 2015 phantom, pipelines recover about 90% of the true bundles but produce four invalid bundles for every valid one, and about half of these "ghost" bundles appear across nearly every group (The challenge of mapping the human connectome based on diffusion tractography). The errors are systematic because every method resolves the same ambiguity, in narrow "bottlenecks" where several fiber populations cross or merge, in the same permissive way.

We propose to treat that ambiguity as a choice made by an adversary rather than as a coin flip, and then compute connectivity exactly. The tracking process becomes a Markov decision process (MDP), a model in which some transitions are random and others are picked by a controller. Inside a single fiber population, movement stays random, reflecting angular spread. At voxels where the local model cannot tell which population a pathway continues into, the controller chooses. Probabilistic model checking is the field of exactly computing probabilities like "reach region B from region A without entering cerebrospinal fluid" in such models. Its tools give, for every region pair, the minimum and maximum connection probability over all ways of resolving the ambiguity. Our central claim is that a connection's minimum probability is a better test of whether it is real than its average streamline count. Ghost bundles should be the ones that exist only under favorable resolutions of bottleneck choices.

Core Claim and What Is New

The research file found no work that applies probabilistic model checking to tractography. Three nearby lines differ in specific ways. Geodesic connectivity mapping casts tracking as optimal control solved by fast marching, running in 7 seconds against 20 minutes for level sets (Control Theory and Fast Marching Techniques for Brain Connectivity Mapping). It is deterministic, uses a single tensor per voxel, and so cannot represent crossings at all. Stochastic completion fields solve a Fokker–Planck advection–diffusion equation between a source and a sink (Probabilistic Anatomical Connectivity Using Completion Fields). This yields a stable connectivity index, but the authors concede that the equation's preference for smooth paths sends flow across adjacent tracts at bottlenecks. Simulation-based inference with neural spline flows gives fast per-voxel posteriors over crossing fibers (Uncertainty mapping and probabilistic tractography using Simulation-based Inference in diffusion MRI). Its global propagation is still ordinary Monte Carlo streamlining.

We reuse the local posteriors from that last line as input. We reuse min/max reachability on MDPs, computed by value iteration or interval iteration, from Automated Verification Techniques for Probabilistic Systems. What none of these works computes is a pair of bounds on connectivity under structured local ambiguity.

The obvious version of this intersection is weaker, and we deliberately move past it. That version models tractography as a plain Markov chain over (voxel, orientation) states and solves for exact reachability. It removes Monte Carlo sampling noise, and nothing more. It computes the same quantity streamlines estimate, so it inherits their structural bias toward ghost bundles. Distance decay also stays, because a long path's probability is a product of many per-step probabilities and that decay is real, not a sampling artifact. We keep the plain chain only as an ablation.

Research Program

State space. States are pairs (white-matter voxel, discretized incoming direction). The direction component lets curvature limits act as transition constraints rather than as post-hoc rules. Successors are neighboring voxels within a forward cone.

Local inputs and the ambiguity rule. For each voxel we take fiber-population estimates from constrained spherical deconvolution, and separately from the Ball & Sticks posteriors produced by simulation-based inference. A voxel becomes a decision point when its posterior assigns substantial mass to two or more populations compatible with the incoming direction. Substantial mass means overlapping angular credible regions, controlled by a threshold $\tau$. At a decision point, each action selects one population. Transitions within that population follow its angular dispersion. Every other voxel keeps purely random transitions. Gray-matter target regions are absorbing goals. Cerebrospinal fluid and out-of-mask exits are absorbing failures.

Main hazard: an overly pessimistic adversary. If every crossing is a decision point, the adversary can divert nearly every path, and the minimum probability collapses to zero everywhere. The threshold $\tau$ guards against this. It limits the adversary to choices the data genuinely cannot rule out, so it governs how informative the bounds are. We sweep $\tau$ and report how the separation between valid and invalid bundles changes with it.

Additional measures. We also report the ratio of minimum to maximum probability. Both bounds decay with path length in a similar way, so the ratio should depend less on length than either bound alone.

Path constraints. These include "avoid cerebrospinal fluid until reaching the target" and waypoint ordering. They enter as products with small automata, as in standard checking of probabilistic temporal-logic formulas.

Computation. One reachability solve per target region gives probabilities from every state at once. For an approximately 2 mm phantom we expect roughly $10^7$–$10^8$ states with about ten successors each. That is within what PRISM reports for structured models (up to $10^{10}$ states). However, PRISM's symbolic engine degrades when numerical values are irregular, and ours will be. We therefore implement sparse interval iteration on GPUs, which gives guaranteed upper and lower error bounds on each computed probability. We validate it against PRISM on downsampled models where PRISM is tractable.

Reachability in MDPs is achieved by memoryless optimal schedulers, meaning the controller's best choice depends only on the current state. The scheduler that achieves the maximum can therefore be read out per state. Read out for a ghost bundle, it serves as a counterexample: it names the specific bottleneck voxels whose resolution creates the connection.

Baselines

Naive reference. The do-nothing point is standard Monte Carlo probabilistic tractography: streamline counts on the same fiber estimates, with matched curvature limits and the same masks. On ISMRM 2015 the reported reference is about 90% of bundles recovered with four invalid bundles per valid one (The challenge of mapping the human connectome based on diffusion tractography). We rerun it ourselves so the comparison uses identical inputs and seeds 10⁶–10⁷ streamlines.

Strongest competitor on real anatomy. On IronTract, the winning pipeline reached a true positive rate of 0.96 at a false positive rate of 0.1 on DSI data. Gaussian tractogram filtering improved accuracy by up to 80% (Insights from the IronTract challenge). We run that style of pipeline, Monte Carlo plus Gaussian filtering, on the same multi-shell data with its filter width tuned on the same held-out injection used to tune our $\tau$. Beating it means a higher partial AUC against tracer ground truth at false positive rates of 0.1 or below. Matching it while also outputting per-connection certificates and counterexamples would be a weaker but still useful result.

Continuous-equation competitor. Stochastic completion fields are the closest sampling-free method. On the Fiber Cup they report indices of 0.17–0.24 for true connections and near zero for unconnected scenarios, while Monte Carlo fell as low as 0.00008 (Probabilistic Anatomical Connectivity Using Completion Fields). We reimplement it on the same grid. Beating it means equal separation on Fiber Cup scenarios d–f and fewer bottleneck jumps on ISMRM.

Ablation: the plain Markov chain. This uses the same states but no decision points, and isolates what exactness alone contributes.

Ablation: the bootstrap. This baseline isolates the adversarial ingredient and is the one that worries us most. We draw K full-brain resolutions of every ambiguous voxel from the posterior, run exact chain reachability on each, and take the 5th percentile. If this matches the MDP minimum, the adversary adds nothing.

No reported result was found on Lacuna for exact or bootstrap reachability on any of these benchmarks.

Why not just…?

Why not just run far more streamlines with anatomical exclusion masks? The ISMRM 2015 study covered 96 pipelines, many with millions of streamlines and anatomical priors, and the 4:1 ratio persisted. Denser sampling estimates the same biased quantity more precisely. The naive baseline directly tests whether our ranking changes which bundles survive, not just how stable their scores are.

Why not just solve a Fokker–Planck equation or use fast marching? These are fast and sampling-free, but they give one number per connection. Fast marching uses a single tensor per voxel and cannot represent crossings. The Fokker–Planck method favors smooth paths, which is exactly the failure at bottlenecks. Neither exposes the ambiguity as something that can be bounded. The ISMRM bottleneck bundles are where this difference should show.

Why not just exact Markov-chain reachability, without the adversary? It is cheaper and removes sampling noise. We expect it to remove noise but not ghosts, and the chain-only ablation will show whether that expectation holds. If the chain alone separates valid from invalid bundles as well as the MDP does, the central claim is wrong, and we would report the work as a fast exact estimator only.

Why not just bootstrap the ambiguous voxels? This is a real threat. Random draws do explore the ambiguity. Our argument is that a ghost bundle needs a coordinated sequence of permissive choices along a path. Independent draws hit that sequence with probability shrinking roughly geometrically in the number of bottleneck voxels, whereas the adversary finds it directly. So the gap between the two should grow with the number of decision points along a bundle. We test exactly this by stratifying bundles by decision-point count and comparing the bootstrap percentile with the MDP minimum within each stratum.

Proposed Experiments and Decisive Checks

The decisive experiment is on the ISMRM 2015 phantom. We rank all candidate region-pair connections by Monte Carlo count, chain reachability, bootstrap 5th percentile, MDP minimum, and the min/max ratio. We then measure the AUC for separating the 25 ground-truth bundles from invalid ones, along with VC, IC, overlap, and overreach after thresholding at matched sensitivity. A second test checks whether the systematic ghost bundles fall disproportionately in the class with low minimum and high maximum probability. A third checks whether the maximizing scheduler localizes them to identifiable bottleneck voxels.

On IronTract we report partial AUC against tracer injections, sweeping $\tau$. On the Fiber Cup we confirm that connected and unconnected scenarios separate at least as cleanly as with completion fields. On the six UK Biobank repeat scans we measure scan-rescan correlation of the minimum and the ratio against Monte Carlo counts.

The proposal fails if the MDP minimum does not beat the bootstrap percentile among bundles with many decision points. It also fails if no value of $\tau$ avoids both a trivial adversary and a collapse of every minimum to zero.