When doctors test how well brain blood vessels can still dilate, they raise the patient's blood carbon dioxide, usually with a short breath-hold, and watch the MRI signal respond. The result is a cerebrovascular reactivity (CVR) map, a per-voxel image of how much each patch of brain can still increase its blood supply and how late it responds. The clinical question such a map is meant to answer is anatomical: which artery's downstream territory has exhausted its reserve? Today that answer usually comes from warping the patient's brain into a standard template and overlaying a fixed textbook map of arterial territories. That atlas assumes one standard arterial layout. The ring of arteries at the base of the brain, the Circle of Willis, has a "complete" textbook form in only about half of people. In a person with a fetal-type posterior cerebral artery, the back of the brain is fed by the carotid artery, not the vertebrobasilar system. A carotid-driven deficit in that person gets reported as a "posterior territory" problem.
We propose to infer a probability distribution over arterial territories for each patient, conditioned on that patient's own vessel graph from angiography (MRA). Delineated regions of impaired reactivity are then forced to respect how blood actually flows down the arterial tree. Two kinds of uncertainty pass through to the final report: expert disagreement about the vessel anatomy, where hypoplastic-versus-absent calls are notoriously unstable, and expert disagreement about where the impaired region ends. The output is a statement a clinician can use, such as "70% probability the delayed region lies in the left internal carotid tree, including its fetal-type posterior branch." It replaces a single deterministic overlay.
Core Claim
Assigning territories to CVR deficits is currently done with a single fixed atlas. That step discards two measurable sources of variation: the subject's arterial variant, and rater ambiguity about both the variant and the deficit boundary. A model that marginalizes over plausible vessel graphs and enforces tree containment should agree better with independent territory measurements and with the spread of expert delineations than either atlas overlay or constraint-free segmentation. The gain should concentrate in patients with non-standard Circle of Willis anatomy. Tree containment here means that an impaired child branch's region lies inside its parent's region, and that a stenosed parent implicates all its descendants.
The central object is a territory posterior. Let $x$ be the subject's CVR amplitude and delay maps, $m$ the subject's MRA, $g$ a candidate Circle of Willis graph (which segments are present, hypoplastic, or absent), and $t_v$ the arterial territory of voxel $v$. Then
$$p(t_v \mid x, m) = \sum_{g} p(t_v \mid x, g)\, p(g \mid m),$$
where $p(g \mid m)$ is a distribution over variants learned from multiple raters rather than a single call. $p(t_v \mid x, g)$ is a variant-conditioned territory prior, obtained from spatially normalized training subjects who share variant $g$, and refined by the subject's own delay map. Impaired-region probabilities $q_a(v)$, one per arterial branch $a$, are trained under the tree constraint $q_{\text{parent}(a)}(v) \ge q_a(v)$ for every voxel.
Relation to Prior Work
Several lines of work motivate the gap. The breath-hold CVR pipeline in Quantitative mapping of cerebrovascular reactivity amplitude and delay with breath-hold BOLD fMRI when end-tidal CO2 quality is low rescues quantitative amplitude and delay maps by predicting end-tidal CO2 from a chest belt. It then judges delayed tissue in Moyamoya disease against a single ground-truth delay map, with no vascular anatomy and no rater model. We take its maps as inputs and ask the territory question it leaves open. Cerebrovascular reactivity measurements using simultaneous 15O-water PET and ASL MRI shows that vasodilation shortens arterial transit time and flags manual arterial delineation as an unchecked error source. That is direct evidence that vessel-level ambiguity leaks into CVR numbers. The 7T small-vessel work of van den Brink et al. and the transcranial Doppler case in De Silva et al. show reserve exhaustion in tissue that looks structurally normal. For this reason the constraint must come from vessels, not from parenchymal lesion masks.
Two challenge papers supply the vessel graph and its instability. TopCoW trains topology-aware segmentation of 13 Circle of Willis components on paired MRA/CTA. We reuse its labels and its centerline-Dice style losses to build $p(g \mid m)$. CROWN reports only 0.26 and 0.27 average balanced accuracy for anterior and posterior variant classification, and explicitly notes inter-annotator disagreement on hypoplasia. We read this as the constraint that forbids a point estimate of $g$.
The hierarchical-constraint literature supplies the method. PrPSeg penalizes violations of subset and exclusion relations between tissue classes, and Anatomy Prior Based U-net penalizes probability mass outside permitted anatomy. Both assume parenchymal nesting against a single consensus label. We replace nesting with arterial feeding order, and the single label with a rater distribution. Rempfler et al. enforce vessel-tree physiology by integer programming on rodent micro-MRA. We borrow the tree logic but apply it to functional maps, as a cheap differentiable penalty rather than a combinatorial solve.
For rater modeling, we reuse the latent-annotator idea of Pionono: each rater is a Gaussian in a latent space, whose mean is their systematic bias and whose spread is their inconsistency. The multi-decoder alternative is OMBA-UNet. Both have been validated on structural or histology images, never on functional territories. Finally, Rathore et al. estimate epistemic uncertainty in vessel segmentation and discard uncertain voxels. We instead keep rater disagreement as signal, and compare the two directly.
Research Program
The work has three components, trained separately so that each can be ablated.
The first is a variant model $p(g \mid m)$. A topology-aware 3D vessel segmenter is trained on TopCoW and CROWN. On top of it, a latent-rater head produces a distribution over present, hypoplastic, and absent labels for each communicating and precommunicating segment. Where only single-reference labels exist, a small in-house panel of three to five neuroradiologists relabels a subset, focusing on the segments CROWN identifies as unstable.
The second is a variant-stratified territory atlas, which is where spatial normalization enters. Rather than one template of territories in standard space, we build one per variant class. This requires subjects with direct territory measurements, collected as part of the new cohort described below with vessel-selective arterial spin labeling. Each subject is normalized on parenchyma, and the territory maps are then averaged within variant class. At test time the atlas is warped back to native space. All tree reasoning happens in native space, which avoids the known failure of parenchymal templates to align individual arteries. The subject's delay map then refines the prior through a simple conditional model: delay is expected to be smooth within a territory and allowed to jump at territory borders.
The third is a tree-constrained, rater-aware impaired-region model. It outputs per-branch impairment maps $q_a$ and is trained with multi-rater delineations of delayed and steal regions under the containment penalty.
Data comes from two sources. For healthy reliability we use EuskalIBUR, which has breath-hold BOLD, end-tidal CO2 and belt traces over 99 sessions in 10 participants. For patients we need a new cohort, which is the main acquisition cost. It would include patients with unilateral carotid or Moyamoya-type steno-occlusive disease, scanned with TOF-MRA, breath-hold BOLD with capnography, and vessel-selective ASL as an independent territory reference. Each impaired region would be delineated by several raters.
Baselines
The naive reference is standard practice. The subject is normalized to MNI space, a fixed textbook territory atlas is overlaid, and the CVR delay map is thresholded as in the breath-hold pipeline of Quantitative mapping of cerebrovascular reactivity amplitude and delay. Territory labels come from the atlas. For delay-region delineation, that paper reports Dice 0.66 against the ground-truth delay map in its Moyamoya case, versus 0.16 with belt-only regressors. We adopt its predicted-CO2 regressor for all methods, so the 0.66 level of delay-map fidelity is the input quality every method shares, not a number to beat. For territory assignment itself, no reported result was found on Lacuna, so all numbers will be established in this project.
The strongest competitor for impaired-region delineation is a consensus-trained 3D segmenter. It is trained on STAPLE-fused multi-rater masks with topology-aware losses as in TopCoW, with uncertainty from an efficient ensemble in the style of Rathore et al.. It is the strongest because QUBIQ found simple ensembles to be hard to beat: top teams reached average Dice of 0.80 to 0.85 per QUBIQ. Rater-specific modeling is represented by OMBA-UNet, which that paper reports outperforming PHiSeg and Probabilistic U-Net in 5 of 7 QUBIQ tasks, and by a Pionono-style latent-rater model without tree constraints. Every competitor gets the same inputs, the same training subjects, the same GPU budget, and an equal hyperparameter search of about 20 configurations. Beating them means lower generalized energy distance to the rater set and higher staged Q-Dice. It also means significantly fewer tree-inconsistent outputs, meaning a child branch flagged impaired outside its parent's region.
Ablations isolate each ingredient. One uses a single hard variant call instead of marginalizing over $g$. One drops the containment penalty. One trains on a single rater. One uses the pooled MNI atlas instead of the variant-stratified one.
Why not just…?
Why not just use the standard territory atlas, since it is what clinicians already use? In subjects with the textbook anatomy it should do about as well as our model, and we expect that. The argument for our model rests entirely on variant carriers. CROWN's 0.26 to 0.27 balanced accuracy implies that variants are common and poorly identified. The decisive test is territory agreement with vessel-selective ASL, stratified by variant class. If the atlas matches our model even in fetal-type posterior cerebral artery carriers, the variant-conditioning component should be dropped.
Why not just cluster the delay map, since delays already trace the feeding territory? This is a real threat, especially in severe steno-occlusive disease, where delays of several seconds may draw territory borders on their own. It is weaker in mild disease and healthy scans. The reported gray-matter delay error is about 0.97 s median, which is comparable to the delay differences between neighboring territories, so borders from delay alone will be noisy there. Our model includes delay as evidence, so the fair comparison is our model against delay clustering at matched disease severity. If clustering wins across the board, the anatomical prior adds nothing and we will report that. We expect it to win only when delays are large.
Why not just fuse readers with STAPLE and train one network? The Latent Doctor Model (paper) found that majority-vote training makes models overconfident on the borderline cases where experts disagree. Hypoplastic segments and watershed borders are exactly such cases. The consensus-trained baseline above tests this directly with calibration and energy-distance metrics.
Why not just use epistemic ensemble uncertainty? Model variance shrinks with more data, but disagreement over whether a vessel is hypoplastic or absent does not. We test this by training on increasing fractions of the data and checking whether ensemble variance still tracks rater disagreement once data is plentiful.
Proposed Experiments
The first check is territory agreement. In the patient cohort, we compare the Dice between each method's most probable territory and the vessel-selective ASL territory, reported separately for standard and variant anatomy. The prediction is a gain concentrated in variant carriers. The second check is agreement with the distribution of expert delineations. Using held-out raters, we measure generalized energy distance and staged Q-Dice following QUBIQ, and count tree-violation rates. The third check is reliability on EuskalIBUR. Territory-level CVR from our model should be at least as reproducible across sessions as atlas-based estimates, which shows the extra flexibility is not fitting noise. The fourth check propagates uncertainty to clinical outputs. For each patient, we report the spread of territory-level CVR induced by sampling vessel graphs $g$ and raters, and test whether cases with high spread are the ones where a blinded clinical panel disagrees on the culprit artery. If that correlation is absent, the uncertainty output is decorative and the claim fails.
References
- Quantitative mapping of cerebrovascular reactivity amplitude and delay with breath-hold BOLD fMRI when end-tidal CO2 quality is low
- Cerebrovascular reactivity measurements using simultaneous 15O-water PET and ASL MRI: Impacts of arterial transit time, labeling efficiency, and hematocrit
- Assessment of Small Vessel Function Using 7T MRI in Patients With Sporadic Cerebral Small Vessel Disease
- Limb-Shaking Transient Ischemic Attack with Distal Micro-Embolic Signals and Impaired Cerebrovascular Reactivity Using Transcranial Doppler
- Benchmarking the CoW with the TopCoW Challenge: Topology-Aware Anatomical Segmentation of the Circle of Willis for CTA and MRA
- Evaluation of techniques for automated classification and artery quantification of the circle of Willis on TOF-MRA images: The CROWN challenge
- PrPSeg: Universal Proposition Learning for Panoramic Renal Pathology Segmentation
- Anatomy Prior Based U-net for Pathology Segmentation with Attention
- Reconstructing cerebrovascular networks under local physiological constraints by integer programming
- Probabilistic Modeling of Inter- and Intra-observer Variability in Medical Image Segmentation
- Inter-Rater Uncertainty Quantification in Medical Image Segmentation via Rater-Specific Bayesian Neural Networks
- Efficient Epistemic Uncertainty Estimation in Cerebrovascular Segmentation
- QUBIQ: Uncertainty Quantification for Biomedical Image Segmentation Challenge
- The Latent Doctor Model for Modeling Inter-Observer Variability