Score-Conditioned Lesion Surrogates for Replicating Stroke Outcome Findings

Most published models that predict lasting disability after stroke from brain scans cannot be checked by anyone else. The cohorts that pair lesion maps with behavioral scores, such as motor or memory test results, are small, single-center and private. The large public scan collections, in turn, have had every clinical score removed to satisfy privacy rules. We propose to train a conditional diffusion model on the paired data, then release that model or samples from it in place of the data. A diffusion model is a generator that learns to turn random noise into realistic samples by gradually denoising it; a conditional one does this while steered by an input. Here the input is a vector of clinical scores and the output is a three-dimensional lesion map in a standard brain template.

We do not claim that synthetic data will make predictions more accurate. We claim something narrower and testable: a surrogate is useful for reproducibility only if published scientific conclusions come out the same when rerun on it. To define "published conclusions" without cherry-picking, we use systematic-review methodology to extract them from the literature. The surrogate is then scored by how many of those extracted findings it reproduces, including the negative ones, such as the finding that depression is essentially unpredictable from lesion anatomy.

Core Claim

A lesion-map diffusion model conditioned on continuous impairment scores can be judged by "claim preservation." This means the fraction of comparative findings from a pre-registered, PRISMA-style review of stroke lesion–outcome studies that are reproduced, with the same direction and similar magnitude, when the analyses are rerun on synthetic data rather than the real cohort. Photorealism, or accuracy gains from augmentation, are not the standard. We expect a diffusion model that is pretrained on many unlabeled lesion masks and then conditioned on scores from a small paired cohort to preserve more claims than two cheaper alternatives, at a measured privacy cost. The first alternative is a simple statistical generator built on lesion components. The second is an unconditional generator whose samples are labeled afterward by a predictor.

The contrast with the closest work is direct. LeFusion paints lesions into real lung CT and cardiac MRI scans. It conditions on masks and intensity histograms, never on behavioral scores, and is judged by segmentation gains. On LIDC it reports nnUNet Dice rising from 78.26% to 83.44%. The multiple sclerosis repainting models, Bi-Directional MS Lesion Filling and Synthesis and MSRepaint, also condition on lesion masks rather than clinical scores, and they address small white-matter plaques rather than large territorial infarcts. Neither line asks whether scientific conclusions survive synthesis. Zu et al. rated all 32 stroke rehabilitation models they reviewed as high or unclear risk of bias. Their main complaints were small samples and pseudo-external validation, but they proposed no data remedy. We use their audit as motivation and their appraisal method as a tool.

Research Program

The first component is the claim registry. Following PRISMA and the CHARMS extraction checklist as Zu et al. did, we will search for studies that relate structural lesion maps to chronic or long-term motor, cognitive or affective outcomes. From each included study we extract every comparative claim that can be re-tested on lesion maps plus scores. Examples are "method A predicts better than method B," "region X carries the signal for domain Y" and "domain Z is not predictable." Each claim gets its reported effect size and a PROBAST risk-of-bias rating, so we can later ask whether low-bias claims replicate more often.

The registry will contain at least the three anchor findings in our file. Forbes et al. found that voxel-wise damage patterns predict motor function far better than regional lesion load (R = 0.83 versus 0.51 for the best corticospinal tract load, in 50 patients). Tozlu et al. used non-negative matrix factorization to break lesion maps into recurring components, and found that left-hemisphere components predict naming and verbal memory while right-hemisphere components predict visuospatial performance. They report R² from 39.7% to 55.8%. The NeuralCup benchmark found that motor outcome is the most predictable domain (R² up to 0.611) while depression is not (R² at most 0.034). The protocol, search strings and extraction rules are registered before any generator is trained, so claims cannot be chosen after we see which ones replicate.

The second component is the generator. We work on lesion masks registered to a standard brain template, not on full MRI. Most registered claims are computed from masks anyway, and masks carry no face or skull, which removes the main identifiability channel that defacing only partly addresses. A reduced-resolution volume also sidesteps much of the memory burden of full 3D generation. We pretrain an unconditional 3D mask diffusion model on the public ATLAS v2.0 training masks (n = 655, drawn from 44 cohorts). These masks have no scores, but they teach the model what plausible stroke shapes look like.

We then add a conditioning branch for scores, in the style of the Volumetric Conditioning Module. It is fine-tuned on paired data, where the condition vector holds domain scores (for example Fugl-Meyer upper extremity, a memory score and a depression scale), age and time since onset. Scores are dropped at random during training so the model also learns to generate without them. At sampling time the strength of score guidance can then be dialed up or down.

That dial is a central object of study. Stronger guidance makes the synthetic lesions more tightly tied to the scores, which could manufacture score–lesion associations that the real data do not support. The conditioning idea of learning control separately from realism follows the decoupling argument in Adding Additional Control to One-Step Diffusion with Joint Distribution Matching. That paper supplies the technique; our question is different.

The third component is privacy auditing. We measure how close each synthetic mask is to its nearest training mask (Dice overlap), and we run membership-inference attacks against the generator. Both are reported next to claim preservation, so that data holders can see the trade-off between faithfulness and privacy rather than a single score.

Baselines

The naive reference point is releasing nothing, together with its opposite, the unattainable upper bound of rerunning each claim on the real data itself. For the upper bound we rerun every registered analysis on the real paired cohort under resampling, which tells us how often a claim replicates even with real data. Many small-cohort claims may fail here. A surrogate cannot be blamed for failing claims that the real data do not support, so preservation is normalized by this real-data rate.

The strongest simple competitor is a statistical generator built on lesion components. It uses the non-negative matrix factorization components of Tozlu et al., fitted on ATLAS masks. We model the joint distribution of component weights and scores with a Gaussian copula, a standard way to build a joint distribution from simple one-variable pieces, fitted on the paired cohort. We then sample and reconstruct masks. It is the strongest competitor because many registered claims are themselves low-dimensional, and a model that is linear in its components may capture them with far less risk of memorizing patients. It receives the same pretraining masks, the same paired cohort, the same number of hyperparameter trials, and the same privacy audit.

The ablation that isolates our key ingredient is an unconditional diffusion model with labels added afterward. It is the same pretrained mask model with no score conditioning. Each sample is assigned scores by a Gaussian process regressor trained on the real cohort, following the design of Forbes et al.. This tests whether learning the joint distribution of lesions and scores matters, or whether a good shape prior plus a predictor suffices. We expect the predictor route to preserve exactly those claims the chosen predictor already encodes, and to break claims that favor other methods.

The strongest reported number for the underlying prediction task is NeuralCup's motor R² of 0.611 on the Brain Before Stroke holdout, with a best multi-domain R² of 0.311 (NeuralCup). No reported result was found on Lacuna for claim preservation or for training on synthetic stroke data and testing on real data. Beating the baselines therefore means two things. First, higher normalized claim preservation than both competitors, at a privacy risk no greater than the copula baseline. Second, models trained only on surrogate data and tested on real holdout patients must keep the NeuralCup ordering across domains. We do not require them to exceed 0.611.

Why not just…?

Why not just release the lesion masks and scores directly? This is a real threat. If a data holder can legally release template-space masks with scores, it dominates any surrogate, and this project is unnecessary for that cohort. The case we target is the common one exemplified by ATLAS v2.0, where even age and sex were stripped under multi-site agreements. It is plausible that unusual lesion shapes combined with scores are linkable to individuals. Our membership-inference audit also measures this risk directly for real released masks, which is useful to data holders regardless of whether the surrogate succeeds.

Why not just use the lesion-component copula? It might suffice. If it preserves as many registered claims as the diffusion model, we will recommend it, because it is cheaper and easier to audit. The diffusion model should win only on claims that depend on fine voxel structure. The best test case is Forbes' finding that voxel patterns beat regional load, since a model built on a few components may compress away exactly the spatial detail that finding rests on. The per-claim breakdown settles this question.

Why not just use federated learning or a hosted benchmark like NeuralCup? Federated training and challenge-style hidden holdouts let people fit models without seeing the data. They do not let an outside group rerun an arbitrary published analysis, such as a new lesion–symptom mapping or a different validation split. That rerunning ability is what Zu et al. found missing. A surrogate complements a hosted holdout rather than replacing it, and we evaluate on the holdout precisely for this reason.

Why not just augment predictors with synthetic data and report the accuracy gain? A generator trained on n patients cannot add information about how scores relate to lesions beyond what those n patients contain. Accuracy gains in the style of LeFusion mostly reflect regularization for segmentation, not new clinical signal. We will report augmentation results for completeness but do not treat them as evidence of validity.

Proposed Experiments

Paired data come from the NeuralCup Brain Before Stroke training release (n = 187), used with nested cross-validation. We will request evaluation against the withheld 50-patient holdout. Unlabeled pretraining uses ATLAS v2.0 training masks, and the ISLES 2022 masks (ISLES 2022) serve as an out-of-distribution shape check. Claims that need full images, such as NeuralCup's FLAIR benefit for cognition, are recorded in the registry but marked as outside the mask-only scope.

The main metric is normalized claim preservation. We also compute Kendall's rank correlation between how registered methods rank on real data and how they rank on surrogate data. The protocol trains on synthetic data and tests on real patients, and it also runs the full analysis on synthetic data alone.

Three checks are decisive. The first is a planted-null test. We shuffle scores in the real training folds before fitting each generator, and every claimed association must then vanish on the surrogate; any surviving association measures signal the generator invented. The second targets the affective domain. The surrogate must reproduce near-zero depression predictability (the real value is at most 0.034), so an R² well above that on synthetic data flags hallucinated structure. We will sweep guidance strength and report the point at which it appears. The third tests whether preservation depends on claim quality. Claims rated low risk of bias by PROBAST should replicate more often than high-risk ones; if they do not, the registry or the generator is uninformative. If the diffusion model fails the planted-null test at usable guidance strengths, or loses to the copula on the per-claim breakdown, we will report that as the finding.