acceptodds
Withdrawn from ICLR 2027

Knowing When Not to Trust: 3D Flow Matching with Self-Flagging Uncertainty for Post-Surgical Cerebral Blood Flow Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Predicting post-operative cerebral blood flow (CBF) from pre-operative imaging could support surgical planning for Moyamoya patients undergoing revascularization. We present two results built on that observation. First, recast as a data-to-data bridge, conditional flow matching clears the identity baseline for the first time, cutting brain-masked MAE 15.7% below it with a third of the sampling steps and beating it for 95–100% of held-out subjects on each of the four brain-masked metrics. Second, the same trained model doubles as its own trust signal: the score of its velocity field is available in closed form, so a Langevin-corrected SDE samples the model's own probability path and yields a per-voxel uncertainty map at no cost in accuracy and with no tuned noise scale. The map orders error monotonically, its most uncertain decile carrying the error of its most confident, and the ordering holds for 40 of 42 held-out subjects. Aggregated to one number per patient it is far weaker, since a subject's error mostly reflects how much that subject changed. We therefore deliver the map rather than a scalar: decision support that shows where a prediction should not be trusted, not an autonomous predictor. All experiments are reproducible and code will be released upon acceptance of paper.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Generative models, Diffusion, Flow matching, Uncertainty estimation, Cerebral blood flow, fMRI, Image regression, 3D images, Brain scans

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_KvyYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes a new method for post surgical cerebral blood flow prediction. They formulate it as a data-to-data conversion [rpblem, and apply conditional flow matching based on the pre-operative images. They then use a Langevin-corrected SDE to sample the model's own probability path and yield a per voxel uncertainty map. Experiments on 3D CBF volumes demonstrate the effectiveness of the proposed method.

Critical strengths
  1. This is a challenging problem with real practical value. Compared to a general image generation task, this work pays attention to the prediction uncertainty which is important in clinical situation.
  2. The uncertainty estimation does not need extra model training. It directly utilizes the velocity field after training to construct the random sampling process. Compared with Deep Ensemble which needs training multiple independent, this work has lower computational cost.
Critical weaknesses
  1. The methodology contribution seems limited. Conditional flow matching is a established technique. The uncertainty estimation method comes from stochastic interpolation theory. The contribution mainly comes from the problem formulation and the systematic integration.
  2. Experiments are not comprehensive. Different models should use the same test split in cross fold validation to ensure a fair comparison. Besides, uncertainty method baselines are not enough. MC dropout, deep ensemble, etc., should be included for comparison.
Questions for authors

Will be 3D CBF volume data be publicly available?

Reviewer_YiVLOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a 3D flow matching framework for predicting post-operative CBF maps using pre-operative images to support surgical planning. It consists of two parts: conditional flow matching for estimating velocity field and stochastic inference and uncertainty estimation. The proposed method reports a 15.7% reduction in brain-masked MAE relative to the identity baseline. Uncertainty is associated with spatial prediction errors, although its utility for identifying unreliable predictions at the patient level remains limited.

Critical strengths

The paper is well-motivated and easy to follow. The proposed approach could provide valuable information on surgical planning for Moyamoya patients. Moreover, the uncertainty map provided along with the predicted post-operative CBF maps to tell a clinician when to trust. These factors are highly crucial in clinic practice.

Critical weaknesses

Limited methodological novelty. The framework primarily adapts conditional flow matching, stochastic interpolants, and sampling-based uncertainty estimation to post-operative CBF prediction. While the application is meaningful, the methodological contribution beyond combining these established components is not clearly demonstrated.

Inconsistent evaluation splits. Different models are evaluated on different held-out subjects, making their relative performance difficult to interpret. Improvements over each model’s identity baseline do not adequately disentangle model performance from differences in cohort difficulty. Although the authors acknowledge the lack of a common test set, this remains a fundamental concern about the validity of the comparisons.

Limited scope of uncertainty assessment. The uncertainty estimates rely on stochastic sampling from a single fixed model and do not explicitly account for epistemic uncertainty. Their ability to identify failures under distribution shifts therefore remains unvalidated. Furthermore, aggregated uncertainty provides little additional information about subject-level prediction error after controlling for actual pre-to-post change (partial correlation = 0.08, p = 0.63). Although these limitations are acknowledged, they restrict the demonstrated scope of the proposed reliability assessment.

Modest error localization utility. In Figure 6, substantial prediction errors extend beyond the highlighted high-uncertainty regions. Although the outline deliberately selects only the most uncertain 10% of voxels, the aggregate result that these voxels capture 15% of total error represents a modest improvement over random selection. The reported spatial association is encouraging, but its practical value for flagging clinically important prediction failures requires further validation.

Questions for authors

Please address the concerns raised in the Weaknesses section.

Additionally, could the authors clarify whether the pre-operative, ground-truth, and predicted images in Figures 4 and 6 share identical grayscale limits?

Reviewer_kGtLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper predicts post-operative cerebral blood flow (CBF) in Moyamoya patients using conditional flow matching (FM). An SDE sampler reuses the velocity field to estimate voxel-wise uncertainty. Reported brain-masked MAE improves 15.7% over identity, with a modest positive correlation between spatial uncertainty and prediction error. However, comparisons do not isolate FM from architecture and task formulation, and the uncertainty ranking needs stronger evidence of added predictive value.

Critical strengths

The clinically motivated task uses a valuable identity baseline: copying the pre-operative scan. This guards against apparent improvements over complex models that still underperform a simple prediction. Estimating sampling uncertainty from the trained velocity field without another network is also worth exploring.

Critical weaknesses

W1. The main models are evaluated on different held-out patients. Subtracting each model’s identity baseline does not remove differences in patient difficulty or splits, weakening comparisons and quantitative interpretation. Differences in patient populations can affect the models tremendously.

W2. A carefully matched direct residual regression baseline is missing. Gains may reflect input preservation, zero initialization, and residual learning rather than the continuous-time FM formulation itself.

W3. The uncertainty claim needs stronger evidence. The final sample variance needs clearer interpretation. Capturing 15% of error in the most uncertain 10% of voxels does not establish superiority over simple practical risk indicators.

W4. The clinical interpretation and some experimental details remain unclear. Separately z-scoring pre- and post-operative scans complicates the interpretation of absolute CBF changes. Missing experimental and reproducibility details further limit clinical claims.

The motivation and observations merit further study, but inconsistent evaluation, a missing simple baseline, and incomplete uncertainty analysis limit the claims. I would reconsider with evidence on common patient splits and stable uncertainty gains over simple indicators.

Questions for authors

Q1. Evaluation on the same held-out patients. Could FM, U-Net, and identity be evaluated on the same patient splits? Baseline subtraction cannot ensure fairness across different surgical changes, registration quality, and difficulty. Please report paired patient-wise errors, means, confidence intervals, and fold variation, and clarify how the 235 pairs, 206 samples, and 42/30 uncertainty-evaluation subjects relate. This would test whether gains persist across comparable patient groups.

Q2. What does FM add beyond direct residual regression? Does Figure 3’s U-Net match FM’s residual output and zero initialization? Please compare direct residual regression with the same backbone, conditioning, and training budget. If a zero-initialized network predicting x_post − x_pre performs similarly, what do the continuous-time bridge and integration add in accuracy, data efficiency, or generalization? This would separate FM’s benefit from the task formulation.

Q3. What uncertainty does the SDE sampler actually measure? Eq. (4) assumes an optimal velocity field. The trained field is only an approximation. Endpoint denoising ideally returns a posterior mean, reducing variance. What does the final standard deviation estimate? Please compare with and without denoising. Please compare ODE and SDE sampling across repeated trajectories with matched initialization, denoising, and cost. Initializing the ODE from the training path’s Gaussian distribution requires no additional arbitrary noise scale.

Q4. Added value of the uncertainty map over simpler measures. Uncertainty correlates similarly with error (0.21) and actual change (0.22). Does it mainly identify larger changes? Please extend the existing coverage analysis with same-patient comparisons against predicted-change magnitude and pre-operative image gradients. These need no repeated sampling. Please also report paired per-patient ranking gains, their variation, and high-error/low-uncertainty cases. What calibration evidence supports clinical reliability claims?

Q5. Interpretation of predictions in standardized CBF space. Separate normalization can remove global perfusion changes. What is the clinical meaning of standardized MAE and related metrics? If possible, evaluate original-scale CBF or vascular territories. Why not condition on surgical approach/side or analyze subgroups? Can standardized predictions assess increased perfusion or hyperperfusion risk? Please qualify clinical decision-support claims accordingly.

Q6. Missing ablations and reproducibility details. Please report raw-change ROI results separately from coherent-change results, and sensitivity to the smoothing scale, ROI fraction, and interpolation noise σ. Clarify checkpoint selection, model training budgets, five-fold aggregation, and validation/test separation. Report full uncertainty-map runtime across trajectories alongside sampling-step efficiency.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.