Hospitals that train a shared brain-tumor segmentation model without pooling patient scans assume that every scan sits in the same anatomical coordinate system. In practice each site warps its scans onto a common brain template, a step called spatial normalization, and that warp can be wrong in ways the site cannot see. Growing tumors push healthy tissue aside and break the template's anatomy. Scanners differ. Some sites have few cases. When a site averages its model with a peer whose scans are badly aligned, it can end up worse than if it had trained alone. Current systems either average everyone or pick partners from a single similarity number. Neither approach asks the question a cautious site should ask: in the worst case consistent with what I know, will joining this group help me?
We propose that each site choose its collaborators by the lower bound of its expected gain, not a point estimate. The site computes this gain from a set of plausible evaluations that explicitly includes plausible alternative registrations of its own validation scans. Sites then reorganize themselves into groups through local join-and-leave moves, with no central server deciding who trains with whom. The central claim can be put in two sentences. Partner-selection errors in federated medical imaging are largely driven by uncertainty about spatial alignment and small validation sets, and a precise similarity score hides both. Replacing that score with an interval whose width comes partly from registration perturbations should sharply reduce the number of sites harmed by collaboration while keeping average accuracy.
What exists and where the gap is
Three bodies of work each supply one piece. Coalition formation, the study of how agents split into groups through local moves, gives us the dynamics. Hedonic games, where an agent's payoff depends only on who is in its own group, let each agent evaluate a move with local information, as in Saad et al.'s task-allocation work. Analyses of single-agent switching (Single-Agent Dynamics in Additively Separable Hedonic Games) state when such moves stop without cycling. The closest machine-learning use is DCFCL. There, decentralized clients form groups each round using a polynomial-time Merge-Blocking algorithm over pairwise similarity, built from gradient alignment and weight distance. That paper motivates our setting but differs from it in two ways. Its "spatial shift" means different label distributions, not different coordinate frames, and its similarity is a precise scalar with no notion of how sure a client is that a peer is compatible.
The only coalition work that reasons under uncertainty is Probability Bounds for Overlapping Coalition Formation. Its agents accept a coalition only if a concentration-inequality lower bound on task success exceeds a confidence level $c$. On social and random graphs this completes more than 2,300 tasks at $c=0.9$, against 637 for expected-value agents. That result motivates our decision rule: acting on a lower bound beat acting on the mean. It also shows what is missing. The bounds are scalar inequalities over assumed independent priors, with no spatial structure and no separation of what is unknown from what is merely noisy.
Imprecise probability supplies the method. A credal set is a set of probability distributions rather than one. The gap between its lower and upper probabilities measures lack of knowledge, as distinct from inherent noise (Hüllermeier et al.). CreDRO builds such sets from ensemble members trained under different robustness budgets and reports 97.4% AUROC on CIFAR10-vs-SVHN against 94.8% for deep ensembles. It is centralized and classification-only. The dossier notes that credal bounds over continuous spatial objects such as deformation fields remain unexplored. We borrow CreDRO's idea of budget-indexed perturbations, but apply it to warps rather than to training distributions.
Spatial normalization supplies the regime and the main source of uncertainty. The BRATS benchmark paper documents that tumor mass effect breaks atlas topology and that expert raters agree only at 74–85% Dice. QU-BraTS found that standard uncertainty maps often flag healthy tissue. Registration-dependent systems treat the warp as exact, including the pelvic few-shot model with a learned affine alignment module and WiCAT's atlas-aligned tokenization, although removing alignment costs the pelvic model at least 12 Dice points. FeTS 2024 runs on atlas-aligned scans and tests collaborator selection, but a central server does the selecting and every quantity is a point estimate. No work in the file has sites choose partners by an interval that accounts for registration error.
Core Claim
Let site $i$ hold a small validation set. Let $S$ be a candidate group, and let $\Delta_i(S)$ be the Dice improvement of the model trained within $S \cup \{i\}$ over $i$'s solo model, measured on $i$'s data. We place a credal set $\mathcal{K}_i(S)$ over the distribution of $\Delta_i(S)$. Its members are indexed jointly by bootstrap resamples of the validation cases and by warps $\phi$ drawn from a plausible-registration set $\Phi_\delta$. The set $\Phi_\delta$ contains the site's nominal atlas warp composed with smooth deformations whose magnitude is at most $\delta$, with larger $\delta$ near tumor and its surrounding region of mass effect. The site's utility is the lower expectation
$$u_i(S) = \min_{P \in \mathcal{K}_i(S)} \mathbb{E}_{P}[\Delta_i(S)],$$
and the solo option has utility zero. Site $i$ moves from its current group $S$ to group $T$ only if $u_i(T) > u_i(S) + \epsilon$. The margin $\epsilon > 0$ guarantees each move improves the mover's lower utility by a fixed amount, which rules out infinite sequences of tiny improvements. Groups that no one wants to leave are Nash stable with respect to lower utilities. Training within each group uses ordinary in-group aggregation.
The prediction is specific. Compared with point-similarity grouping and with averaging everyone, the cautious rule should cut the negative-transfer rate, the fraction of sites whose final Dice is below their solo Dice. It should do so mostly by excluding sites with large registration error, and it should cost little in mean Dice.
Research Program
We build the system in three layers.
The first layer computes the plausible-warp set. Each site perturbs its own validation images and labels with smooth random displacement fields, bounded in magnitude by $\delta$. It sets $\delta$ locally from registration-quality signals it can compute without labels: inverse-consistency error of its registration and the image-similarity residual after warping, plus a larger budget inside a dilated tumor mask. This is the spatial-credal ingredient, and it never leaves the site.
The second layer is the coalition dynamics. Following the proposer-sampling idea of the probability-bounds paper, each round a site evaluates only a few candidate groups. It does this by downloading their current in-group model, which is standard in personalized federated learning, and computing $u_i$. Moves are made asynchronously, so no server is needed.
The third layer is analysis. Lower expectations do not make the game additively separable, so the existing convergence results do not transfer directly. We will prove termination from the $\epsilon$ margin together with bounded utilities. We will also measure empirically how many moves occur and whether cycling happens when the margin is removed.
Baselines
The naive reference point is local training only: each site trains on its own data and never collaborates. By construction this has zero negative transfer, so it sets the floor any collaborative method must clear on mean Dice.
The strongest existing method is FedPOD, the PID-stabilized aggregation that won FeTS 2024. It reports mean Dice of 0.733 (enhancing tumor), 0.761 (tumor core) and 0.751 (whole tumor) on the FeTS 2024 partitions, with a convergence score of 0.764. We will rerun it on the same 23 real-site and 33 artificial-site partitions with the challenge's round budget. We also use FedPOD as the in-group aggregator inside our coalitions, so any difference comes from who trains with whom, not from how weights are averaged. Beating it means at least halving the negative-transfer rate while staying within one standard error of its mean Dice. A gain in worst-site Dice would be a bonus, not a requirement.
The strongest partner-selection competitor is DCFCL-style grouping by precise similarity (DCFCL), run with Merge-Blocking over gradient alignment and weight distance. A second version uses the point estimate $\mathbb{E}[\Delta_i(S)]$ under the nominal warp. This second version matches the expected-value agents that the probability-bounds paper beat in a non-spatial setting. No reported FeTS result exists for either, since no reported result was found on Lacuna for coalition-based selection on FeTS. We will run both, giving every method the same number of evaluations per round and the same tuning budget for its threshold or margin.
Two ablations isolate the key ingredient. The first keeps the credal rule but sets $\delta = 0$, so the interval reflects only bootstrap and training variability. The second replaces the credal lower expectation with a Gaussian lower confidence bound, mean minus $k$ standard deviations, computed from the same evaluations. The $k$ value is tuned on the same budget as our $c$ and $\delta$ scale.
Why not just…?
Why not just average everyone with a robust aggregator like FedPOD? Robust aggregation damps noisy updates. It still forces every site to accept the shared model, and FeTS reports only mean Dice, which can hide sites that collaboration hurts. Our per-site negative-transfer measurement will show directly whether averaging harms the badly registered sites. If it does not, there is little room for any selection method, and we would report that.
Why not just fix the registration? Better registration helps, but the BRATS paper shows that mass effect invalidates the atlas topology itself. Some misalignment is therefore irreducible rather than an engineering bug. We will test this by running all methods after an improved nonrigid re-registration. If the benefit of the cautious rule disappears there, the problem was fixable upstream.
Why not just use mean-minus-standard-deviation instead of credal sets? This is the most serious threat. A lower confidence bound over the same perturbed evaluations may behave almost identically. The second ablation in the Baselines section settles it. If the Gaussian bound matches the credal rule, we will drop the credal machinery and claim only that registration-perturbed lower bounds matter. We expect a difference in one specific case: warp perturbations produce skewed, bimodal gain distributions near tumors, where a minimum over the set and a variance-based bound diverge.
Proposed Experiments
The main testbed is FeTS 2024 multi-parametric MRI. We report Dice and 95th-percentile Hausdorff distance (HD95) per tumor compartment, negative-transfer rate, worst-site Dice, and the number of coalition moves.
The decisive check is a controlled corruption study. On a known subset of sites, we inject misregistration of graded magnitude using random affine plus smooth nonrigid warps. Because we know which sites are corrupted, we can measure how precisely and how completely each selection method excludes them, and whether the registration term of the interval grows with the injected error. Our hypothesis fails if point-similarity grouping excludes corrupted sites as well as the credal rule does. It also fails if the $\delta = 0$ ablation performs as well as the full rule.
A secondary check uses the six-institution pelvic MRI setting from the few-shot pelvic paper, which reports 41.73% Dice on novel classes. In that anatomy, affine alignment is known to be insufficient, so it tests whether the effect extends beyond brain. We also use QU-BraTS filtered-true-positive and filtered-true-negative ratios to check that the warp-derived widths concentrate near pathology rather than in healthy tissue, the failure QU-BraTS found in precise uncertainty maps.
References
- Hedonic Coalition Formation for Distributed Task Allocation among Wireless Agents
- Single-Agent Dynamics in Additively Separable Hedonic Games
- Decentralized Dynamic Cooperation of Personalized Models for Federated Continual Learning
- Probability Bounds for Overlapping Coalition Formation
- Quantification of Credal Uncertainty in Machine Learning: A Critical Analysis and Empirical Comparison
- Learning Credal Ensembles via Distributionally Robust Optimization
- The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS)
- QU-BraTS: MICCAI BraTS 2020 Challenge on Quantifying Uncertainty in Brain Tumor Segmentation – Analysis of Ranking Scores and Benchmarking Results
- FEW-SHOT Image Segmentation for Cross-Institution Male Pelvic Organs Using Registration-Assisted Prototypical Learning
- Cross-Subject Modeling for Widefield Calcium Imaging via Atlas-Aligned Spatiotemporal Tokenization
- The MICCAI Federated Tumor Segmentation (FeTS) Challenge 2024: Efficient and Robust Aggregation Methods