Letting Segmentation Uncertainty Choose the Atlas Resolution for Cross-Site Alzheimer's Staging

Brain MRI studies summarize a scan by cutting the brain into named regions, an atlas or "parcellation," and measuring each one. How finely to cut is an arbitrary choice, and results depend on it: diagnostic accuracy can swing by more than ten points between atlases. Fine cuts are also the first thing to break when a scan comes from an unfamiliar hospital scanner, because the boundaries between small neighbouring regions become ambiguous. We propose that the atlas resolution should not be fixed in advance. A 3D segmentation network that labels regions along a coarse-to-fine anatomical tree already knows, scan by scan and region by region, how sure it is about each split. We use that confidence to decide where each scan's measurements stop. Where boundaries are clear, the scan is described by fine regions. Where they are not, because of anatomy, atrophy, or an unfamiliar scanner, it falls back to the parent region.

The question is whether this per-scan choice of resolution makes fine-grained disease staging more robust when the model moves to a new site. Fine-grained staging here means separating stable from progressive mild cognitive impairment (MCI), not just healthy from diseased. Staging is where the stakes are highest and labels are scarcest. It is also where small, unreliable regional differences are most tempting to overfit. Domain shift supplies the stress test, parcellation sensitivity is the problem being fixed, and native 3D context is what makes the confidence signal trustworthy enough to use.

Core Claim

The claim is that scanner shift reaches regional brain measurements mostly as boundary ambiguity between sibling regions. A hierarchical 3D segmenter's branch-level confidence can detect this ambiguity at test time without labels. Pooling regional features up the tree in proportion to that uncertainty should therefore give staging features that are both less site-dependent and less atlas-dependent than any fixed resolution, while keeping fine detail where it is reliable.

This differs from the nearest work in a specific way. Hierarchical Brain Parcellation with Uncertainty built a native 3D UNet on ADNI2 T1 scans. It predicts branching decisions over a 6-level, 151-leaf anatomical tree and shows that uncertain voxels can "degrade gracefully" to coarser labels. However, it stops at segmentation: it never tests a new scanner and never feeds the uncertainty into any diagnostic model. Multi-Atlas Brain Network Classification through Consistency Distillation and Complementary Information Fusion (AIDFusion) addresses atlas sensitivity by fusing two fixed atlases with consistency losses. Its atlases are static, its inputs are pre-extracted ROI time series, and its tasks are binary. Our proposal reuses the first paper's mechanism and asks a question neither paper asks: whether label-free, per-scan resolution selection stabilizes ordinal staging across sites.

Mechanism

A native 3D hierarchical segmenter, trained as in the hierarchical parcellation paper on the GIF label tree, gives each scan soft probabilistic labels at every tree node. It also gives a confidence $c_v \in [0,1]$ for the decision that splits node $v$ into its children. We compute $c_v$ as one minus the normalized entropy of that branching decision, averaged over voxels within a few millimetres of the internal boundaries of $v$, and set $c_\ell = 0$ at leaves.

For every node we compute a morphometric score $z_v$: soft volume, corrected for intracranial volume and age, then z-scored against cognitively normal training subjects. This makes parent and child scores comparable. Each leaf $\ell$ then receives a stopping-depth-weighted score:

$$\tilde z_\ell = \sum_{v \in \mathrm{path}(\ell)} \pi_v \, z_v, \qquad \pi_v = (1 - c_v) \prod_{a \in \mathrm{anc}(v)} c_a ,$$

Here $\mathrm{path}(\ell)$ is the chain of nodes from the root down to $\ell$, and $\mathrm{anc}(v)$ is the set of strict ancestors of $v$. The weights $\pi_v$ act as a stick-breaking distribution over where to stop descending, and they sum to one because $c_\ell = 0$. A confidently split region contributes its own fine score. An ambiguous one inherits its parent's score. The 151-dimensional vector $\tilde z$ has the same length for every scan, so a downstream staging head never sees a changing input space.

The staging head is a small network trained with the hybrid-granularity ordinal loss of HOPE over the ordered stages normal → stable MCI → progressive MCI → AD. That loss combines instance-to-instance ranking, instance-to-prototype compactness, and prototype separation. We reuse it unchanged, so any gain comes from the features, not the loss.

Research Program

The work has three parts. First, we test the premise directly. We measure whether branch confidence $c_v$ actually drops on scans from unseen sites (NACC, AIBL) relative to held-out ADNI scans, and whether those drops line up with the regions whose fixed-resolution volumes carry the most site variance. The original hierarchical paper never evaluated this out of distribution, so it is a real open question. Segmenters are often overconfident under shift, and if that holds here the mechanism has no signal. We will therefore calibrate confidences on ADNI with temperature scaling and report calibration error on external sites.

Second, we test the volumetric-context ingredient. We train a slice-wise 2D version of the same hierarchical segmenter and compare it with the native 3D one. The prediction is that the 2D version's uncertainty will be higher and noisier in the out-of-plane direction, which would make it a worse gate. This prediction draws on Deep convolutional networks for pancreas segmentation in CT imaging, which found that ambiguous boundaries resolve only along the depth axis, and on the observation in Towards Generalisable Foundation Models for 3D Brain MRI that its own pseudo-3D slice stacking lacks full volumetric attention.

Third, we run the staging comparison. Its motivating evidence is the 13.3% atlas-induced accuracy swing reported in NF-GAT. A second motivation is the finding in Intra-Atlas Node Size Effects on Graph Metrics that region size interacts with disease status in ADNI. If size and disease interact, a fixed resolution confounds the two, while a per-scan resolution at least makes the choice explicit and data-driven. The training protocol matches HOPE's: train on ADNI, then test on held-out ADNI and on the external NACC cohort, with AIBL as a second external site.

Baselines

The naive reference fixes the atlas at the finest leaf level, feeding raw $z_\ell$ to the identical HOPE-loss head. A second naive reference fixes it at a single coarser level chosen on ADNI validation. Together they bracket the "pick one atlas" practice and show how much atlas choice alone moves staging accuracy within our pipeline.

The strongest feature-level competitor concatenates every tree level ($z_v$ for all nodes) and lets the head learn a static weighting. This baseline isolates the key ingredient. It has access to exactly the same information except the per-scan confidence. If it matches our method, the uncertainty gating is unnecessary.

We also run a gating ablation that uses the same stick-breaking pooling with confidences fixed at their ADNI population means, and another that uses confidences from the 2D segmenter. These separate "multi-resolution pooling" from "per-scan adaptive pooling" and from "3D-derived confidence."

The standard harmonization competitor applies neuroCombat to regional features. Evaluation of Mean Shift, ComBat, and CycleGAN showed it removes site effects on first-order connectivity summaries but fails on path-dependent ones. ComBat requires a batch of unlabeled target-site scans, which our method does not. We therefore report it in two forms: alone, with that extra access disclosed, and combined with our features, to test whether the two are complementary.

The strongest end-to-end competitor is HOPE itself, a 3D CNN on whole T1 volumes. It reports 72.0% accuracy for MCI progression on the ADNI internal test and 67.6% on external NACC, outperforming 18 baselines. We will rerun it with its published loss on our exact subject splits so that both methods see the same data and the same tuning budget, with the same number of hyperparameter trials on ADNI validation. For coarse AD-versus-normal detection we also report BrainFound, which reports AUROC 0.883 on NACC against 0.865 for supervised baselines. No reported result was found on Lacuna for progression staging on AIBL by any method.

Beating the field means three things. First, a NACC progression accuracy above the rerun HOPE figure, while ADNI accuracy is not lower. Second, a smaller ADNI-to-NACC drop than HOPE's 4.4 points. Third, a smaller accuracy spread across fixed resolutions than the naive references show, with every comparison carrying bootstrap confidence intervals over subjects.

Why not just…?

Why not just run HOPE end-to-end on the raw volume and skip parcellation? This is the real threat. HOPE already handles granular staging and cross-site transfer from 3D T1 images, and an end-to-end model has no atlas to be sensitive to. Our answer is empirical rather than rhetorical. If the matched HOPE rerun equals or beats our method on NACC and AIBL, the staging-accuracy claim fails. The contribution then shrinks to a regional, site-stable morphometric representation whose resolution choice is auditable. We report site variance explained per region for both approaches, using HOPE's internal features probed regionally, so that the stability claim can be judged separately from the accuracy claim.

Why not just apply neuroCombat to fixed-atlas features? ComBat models site as additive and multiplicative shifts per feature. Boundary misassignment between sibling regions is different: volume moves from one child to its neighbour, which per-feature rescaling cannot undo. The harmonization benchmark's failure on betweenness centrality is consistent with this, since that metric is sensitive to exactly such local redistributions. ComBat also needs target-site batches and site labels. The ComBat-alone versus ComBat-plus-ours comparison settles whether the two address the same error.

Why not just concatenate all resolutions, or fuse multiple atlases as AIDFusion does? Both learn one weighting at training time on ADNI and cannot react to a scan whose boundaries are ambiguous only because it came from an unseen scanner. AIDFusion's reported gains, including 6.77% over the best baseline on Matai, show that fusion helps within its benchmarks. It was not evaluated on ordinal staging or on held-out sites. The all-levels concatenation baseline and the population-mean gating ablation are designed to show whether per-scan adaptivity matters. If they tie our method, we will report that the adaptivity is unnecessary.

Decisive Checks

The project lives or dies on three results, each of which could come out negative. Calibrated branch confidence must fall measurably on NACC and AIBL, and specifically in the regions whose fixed-level volumes show the most site variance; if not, the gate has nothing to read. The uncertainty-gated features must beat both the all-levels concatenation and the population-mean gate on external progression accuracy; if not, the idea reduces to multi-scale pooling. The 3D-derived gate must outperform the 2D-derived gate; if not, volumetric context is not doing the work we claim. If all three hold and the method also matches or beats the rerun HOPE on NACC, we will have shown that choosing atlas resolution per scan from the segmenter's own doubt is a practical answer to both parcellation sensitivity and scanner shift in fine-grained staging.