acceptodds
Withdrawn from ICLR 2027

EviFlow: Structured-Noise Flow Matching for Multi-Rater Segmentation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-rater segmentation requires coherent alternative delineations, not merely variable pixelwise predictions. We introduce EviFlow, a conditional flow-matching sampler driven by spatially coherent base noise, together with a distribution-aware evaluation protocol. We establish that, for binary masks on a fixed finite grid, the generalized energy distance (GED) with ground distance is exactly the squared maximum mean discrepancy under the Jaccard kernel. Its strict positive definiteness makes this GED a strictly proper population divergence; the variant with ground distance is strictly proper as well. We distinguish these population guarantees from U- and V-statistic estimation and complement GED with Hungarian-matched IoU (HM-IoU), dispersion, and simple stochastic references. Replacing independent base noise with normalized blurred Gaussian noise yields coherent whole-object alternatives: on LIDC-IDRI with four raters and nine seeds, energy GED decreases from 0.178±0.014 to 0.092±0.022 and HM-IoU rises from 0.361±0.027 to 0.489±0.029. Baseline tuning changes the comparison substantially: a validation-tuned Probabilistic U-Net is not significantly different on energy GED and is better on Jaccard MMD and HM-IoU, whereas its untuned counterpart is the weakest learned stochastic baseline on LIDC GED. On LNDb, EviFlow achieves lower energy GED than the tuned Probabilistic U-Net; additional benchmarks reveal dataset-dependent trade-offs. These findings show the value of structured generative noise and establish a practical case for reporting distributional fit, sample fidelity, dispersion, and baseline tuning together.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-rater segmentation, Flow matching, Structured noise, Generalized energy distance, Maximum mean discrepancy, Uncertainty quantification

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_NRVeOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposed a conditional flow-matching sampler driven by spatially coherent base noise, together with a distribution-aware evaluation protocol for multi-rater segmentation. Some ablations are provided to demonstrated the designs.

Critical strengths

None

Critical weaknesses
  • Limited novelty against existing works The main contribution is a fixed smoothing and normalization of the initial noise. Section 2 primarily lists relevant approaches without analyzing their limitations or explaining why the proposed construction addresses an unresolved problem. For example, the distinction from Flow-SSN is largely described as using a fixed rather than learned base distribution, without sufficiently investigating the implications of this choice. It is recommended to include more discussion on empirical evidence.

  • Insufficient comparison with recent competitive methods The main comparison approaches are MC-drop, U-Net, and SSN, which are not the latest diffusion- or flow-based methods. Other directly relevant methods cited by the paper are not experimentally compared. The limited coverage of recent methods makes it difficult to assess whether Eviflow advances the current state-of-the-art.

  • Main ablation is incomplete Independent noise without the hinge is missing in Table 3. The used framework in Figure 3 uses a 3x3 convolution and two residual blocks. It is recommended to test on larger-receptive-field.

  • The method presentation lacks clarity and sufficient implementation detail An architecture diagram is available in Appendix C, but the main text lacks an integrated overview that guides the reader through the method. The necessity of the evidential formulation is also unclear given the similar HM-IoU reported for the softmax-head alternative. Several mathematical details require clarification. The total objective includes losses whose exact formulations are not clearly specified in the method description, and the diversity hinge used in the central ablation is not explicitly defined. In Eq. (9), the linear transformation producing q does not itself guarantee that q lies strictly within (0,1), as required for the stated Bernoulli KL. Any clipping, bounded parameterization, or other safeguard should be specified. These omissions affect understanding and reproducibility, beyond presentation quality alone.

  • Limited theoretical contribution The GED characterization combines the established energy–MMD correspondence with the established strict positive definiteness of the Jaccard kernel on finite set spaces. However, the theory concerns the interpretation of an evaluation metric and does not explain why the proposed structured-noise sampler should perform better.

Questions for authors
  1. Why are other recent, directly relevant flow- and diffusion-based segmentation methods omitted from the comparison?

  2. What are the results for independent noise without the diversity hinge under otherwise identical training settings?

  3. Does the benefit of correlated noise persist with larger-receptive-field velocity networks and across different blur scales?

Reviewer_Dz93OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes EviFlow for multi-rater segmentation using flow matching with spatially correlated Gaussian noise. It also discusses the connection between GED and MMD and examines how evaluation metrics and baseline tuning affect the comparison of probabilistic segmentation methods.

Critical strengths

The paper is clearly written and the experiments are fairly comprehensive. I appreciate the careful evaluation of different metrics and the inclusion of a properly tuned Probabilistic U-Net baseline. The analysis of finite-sample GED estimation is also useful.

Critical weaknesses
  1. My main concern is the limited novelty. Flow matching has already been applied to probabilistic segmentation, and spatially correlated priors have been explored in previous work, including Flow-SSN. The proposed method mainly replaces independent Gaussian noise with a fixed smoothed Gaussian field. I am not convinced this represents a sufficiently new methodological contribution.

  2. The theoretical contribution also seems rather incremental. The relationship between energy distance and MMD, as well as the positive definiteness of the Jaccard kernel, are established results. Moreover, EviFlow does not outperform the tuned Probabilistic U-Net on the main LIDC benchmark. This makes the practical advantage of the proposed method less convincing.

Questions for authors
  1. What is the main methodological novelty compared with Flow-SSN and previous work using correlated priors in generative models?

  2. Which theoretical results are genuinely new beyond applying existing GED, MMD, and Jaccard kernel results to segmentation?

  3. Why are some cross-dataset comparisons based on baselines that are not fully tuned, given that baseline tuning is one of the paper's main messages?

Reviewer_tgZiOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The EviFlow model simulates segmentation uncertainty through conditional flow matching driven by spatially correlated noise, thereby generating coherent segmentation mask samples. In addition to this model, this paper theoretically analyzes the generalized energy distance (GED) and develops an evaluation protocol for probabilistic segmentation. Experimental results on the LIDC-IDRI, LNDb, QUBIQ, and Cityscapes datasets show that structured noise helps improve mask quality and enhance consistency with the reference distribution. However, a carefully tuned probabilistic U-Net matches or even surpasses EviFlow on several metrics, indicating that the model's performance largely depends on baseline optimization and evaluation conditions.

Critical strengths

This evaluation method combines GED, HM-IoU, dispersion indices, and statistical tests to assess multiple aspects of probability segmentation. Attention to baseline optimization is also crucial, as it demonstrates how optimal choices influence comparison conclusions and encourages more reliable benchmarking.

Critical weaknesses
  1. The study found that parameter tuning significantly altered model rankings on the LIDC-IDRI dataset, raising concerns about the consistency of other comparisons. Specifically, the QUBIQ and Cityscapes experiments did not demonstrate that the competing methods achieved a comparable amount of optimization work. Without this information, it is difficult to determine whether the reported advantages reflect stronger generalization ability or differences in baseline preparation.

  2. Comparisons with independent noise support the selection of spatially correlated noise in EviFlow, but comparisons with well-tuned competing methods fail to provide sufficiently consistent evidence of an overall advantage. The paper should clearly indicate in which settings or real-world applications EviFlow demonstrates a significant advantage over existing probabilistic segmentation methods.

Questions for authors
  1. How does improved distribution consistency translate into useful clinical outcomes? For example, can sampling masks provide more reliable estimates of lesion volume uncertainty or support diagnostic decisions? Evidence from downstream tasks will help determine whether characterizing inter-rater variability is more advantageous than a single deterministic prediction.

  2. What impact does the fuzz kernel size have on segmentation quality and distribution performance? Since all reported experiments used a size of 17, explain whether an appropriate noise correlation scale varies with image resolution or lesion size.

  3. Given that the HM-IoU values ​​for softmax and the evidence head are nearly equal (0.488 and 0.489, respectively), what measurable advantages does the evidence approach offer? Are these advantages manifested in uncertainty calibration, other distribution metrics, or specific use cases not covered by HM-IoU?

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

Related papers

Discussion (0)

Sign in to comment.