acceptodds
Under review as a conference paper at ICLR 2027

Uncertainty-Guided Spectral Adaptation for Coronary Artery Segmentation

Abstract

Coronary artery segmentation in invasive coronary angiography (ICA) is hindered by changes in imaging conditions and ambiguous vessel boundaries. These challenges are particularly pronounced for thin, low-contrast branches, whose structural details are easily lost and whose uncertain pixels can compromise cross-domain alignment. We propose FUP-Net, a Frequency-aware Uncertainty-guided Prototype Network that couples spectral representation learning with uncertainty-guided adaptation. Its Frequency-aware Heterogeneous Mamba2 (FH-Mamba2) module captures complementary frequency information by combining long-range vascular modeling with convolutional processing of local details. Building on these features, an Uncertainty-guided Prototype Adaptation (UPA) mechanism reduces the influence of ambiguous pixels when updating semantic prototypes, enabling more reliable alignment across imaging domains. Experiments on multiple ICA benchmarks assess segmentation accuracy, vascular topology, and cross-domain generalization. FUP-Net improves Dice by 2.28, 7.40, and 3.65 percentage points over the strongest baselines on JSPH-ICA, DCA1, and ARCADE, respectively. Ablation studies further validate the contributions of frequency-aware modeling and uncertainty-guided prototype adaptation.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Representation Learning, Medical Image Segmentation, State Space Models, Mamba, Domain Adaptation, Uncertainty Quantification, Prototype Learning, Frequency-aware Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_VideOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes FUP-Net for coronary artery segmentation. Its FH-Mamba2 module separates smooth and detail features, using Mamba2 for long-range structure and convolutions for local vessel details. Its UPA module uses prediction uncertainty to weight target-domain features when updating semantic prototypes, followed by conditional domain alignment. The paper reports results on JSPH-ICA, DCA1, and ARCADE, including segmentation, boundary, cross-domain, and component-ablation evaluations.

Critical strengths
  1. The frequency-aware component has a useful ablation. Adding frequency separation to the Mamba2 variant improves Dice by 2.19, 3.24, and 2.11 percentage points across the three datasets (Table 3). This provides evidence for processing smooth vessel structure and local detail with different operators.

  2. The evaluation reports results across three ICA datasets using both region-overlap and boundary metrics. This breadth helps assess whether the reported improvements extend beyond Dice alone, subject to clarification of the training protocol below.

Critical weaknesses
  1. The supplied implementation appears inconsistent with the paper’s central unlabeled-target adaptation claim. Section 4.2 and Eq. (15) state that target-domain images are used without segmentation labels. Yet the supplied data.py requires target training masks, and train.py computes a segmentation loss using those masks and adds it to the total objective (loss_t, lines 297–330). Appendix C also describes experiments without UPA as supervised training “on each dataset,” which conflicts with Section 4.2’s source-only description. If the supplied training code generated the DCA1 and ARCADE results, their gains cannot be interpreted as evidence for adaptation using unlabeled target data. This uncertainty directly affects the paper’s main empirical claim.

  2. The contribution of uncertainty weighting is not sufficiently isolated. In Table 3, UPP+UPA improves over UPP+ProtoDA by 0.62 Dice points on DCA1 and 1.17 on ARCADE, without uncertainty estimates for these differences. The full-model comparison changes multiple adaptation components at once. Moreover, Table 1 lacks adaptation baselines with clearly matched access to target-domain data, limiting attribution of the reported cross-domain advantage.

Questions for authors
  1. Were target-domain ground-truth segmentation masks used to train any model reported in Table 1? Please identify the exact code version and configuration used for those results and explain why the supplied training code includes loss_t computed from target masks. If the reported models used target labels, please state that setting explicitly; if they did not, please provide the corresponding unlabeled-target training code.

  2. For each Table 1 method, what labeled source data and target data were available during training and model selection? Please provide comparisons against adaptation baselines with the same data access.

  3. With FH-Mamba2 and the other adaptation components held fixed, what is the effect of removing only the uncertainty weighting from UPA? Please report variability or paired uncertainty for this ablation.

Reviewer_os9FOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FUP-Net combines frequency-aware feature processing with prototype-based domain adaptation for coronary artery segmentation. It sends smoothed features to Mamba2 and residual features to a CNN. Prediction entropy controls how target features update the prototypes. The paper reports results on three angiography datasets.

Critical strengths
  1. The method reports higher Dice scores than the compared models on all three datasets. The results suggest practical value for coronary artery segmentation.
  2. The paper includes component ablations and computational costs. These help assess the gains and the additional overhead.
Critical weaknesses
  1. Low frequency is not clearly linked to vessel topology. A smooth vessel trajectory does not imply low-frequency features. Thin connections can contain high-frequency details yet determine connectivity. The paper does not show what anatomical information the smoothed feature maps actually preserve.
  2. The choice of Mamba2 for the low-frequency branch is weakly justified. Frequency and dependency range are different properties. Fine branches may also need distant context. Appendix A does not establish the proposed branch assignment for selective Mamba2. A matched comparison with reversed assignments is missing.
  3. The benefit of uncertainty weighting for difficult branches is unclear. Faint branches may receive low weights, while confidently missed branches may still receive high weights as background. The paper does not show how UPA handles these cases. Its uncertainty component is not isolated on the full FH-Mamba2 backbone.
  4. The topology claims lack direct quantitative support. The abstract mentions topology evaluation, but the reported metrics measure overlap, sensitivity, specificity, and boundary distances. None directly measures connectivity. A centerline or connectivity metric would better support the claimed structural improvements.
  5. The training protocol is inconsistent. Section 4.2 describes source-only training for FUP-Net without UPA. Appendix C instead describes supervised training on each dataset. These are different settings, so it is unclear how the target-domain results in Table 1 should be interpreted.
  6. The domain adaptation comparison is incomplete. Table 1 mainly compares segmentation networks. It does not establish whether they have the same access to unlabeled target data. The internal DA and ProtoDA ablations are useful, but do not replace a matched comparison with an established adaptation method.
Questions for authors
  1. What evidence shows that the smoothed feature component represents vessel topology? Why should Mamba2 process this component rather than the detail component? A matched comparison with reversed branch assignments would help test this assumption.
  2. Does uncertainty weighting improve recall on faint or thin branches? How does it handle branches that are confidently misclassified as background? Please isolate its effect on the full FH-Mamba2 backbone.
  3. Please clarify the training protocol. Section 4.2 describes source-only training for FUP-Net without UPA, but Appendix C describes supervised training on each dataset. Which setting produced Table 1? Did the competing methods have the same access to unlabeled target data?
Reviewer_Lao2OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes FUP-Net for coronary artery segmentation in ICA. It combines a UNet++ backbone with frequency-aware Mamba2/CNN branches and uncertainty-guided prototype adaptation. Experiments on JSPH-ICA, DCA1 and ARCADE report segmentation improvements, supported by component ablations, computational comparisons and significance tests.

Critical strengths

(1) I find the problem meaningful. Preserving thin vessels while adapting across angiography datasets is a useful goal, and combining local detail with longer-range context is a reasonable design direction. (2) The within-domain result is encouraging: FH-Mamba2 improves JSPH-ICA Dice from 85.57 to 89.58 without adaptation. I also appreciate the progressive DA, ProtoDA and UPA ablations, rather than reporting only the final model.

Critical weaknesses

(1) The cross-domain training protocol is contradictory. My biggest concern is that I cannot tell what the main comparison actually measures. Section 4.2 says FUP-Net without UPA uses source supervision only, while Appendix C says it is trained with supervised learning on each dataset. The target-domain training protocol of the seven baselines is also unclear. These are not equivalent settings. Until this is resolved, I cannot attribute the reported cross-domain gains to the proposed method. (2) Data independence and model selection are insufficiently documented. JSPH-ICA contains 314 images from 99 patients, but patient-disjoint splits are not stated. Appendix C selects checkpoints by validation Dice without identifying which domain supplies the labels during adaptation. Target test-image access is also unspecified. Additionally, the aggregation of five fold models for the per-image Wilcoxon tests is unclear. I am not claiming leakage, but these omissions prevent me from judging the independence of the evaluation and its reliance on target supervision. (3) Limited methodological novelty. I find the methodological novelty limited. FH-Mamba2 combines a pooling-based smooth/residual split with Mamba2 and CNN branches, while UPA combines entropy weighting, class prototypes and conditional adversarial alignment. Related ideas already appear in SFD-Mamba2Net (Mu et al., 2025), CDAN+E (Long et al., NeurIPS 2018) and ProDA (Zhang et al., CVPR 2021). The exact implementations differ, but I mainly see a task-specific integration of existing ideas. I do not see a sufficiently clear new mechanism or insight from coupling these two modules. The current ablations show component gains, but do not establish why this particular combination is a substantial methodological advance. (4) I am not convinced that uncertainty weighting helps the difficult vessels motivating the paper. Equation 9 assigns weights using relative entropy within a batch, which is not itself evidence of prediction correctness. My concern is whether it suppresses wrong assignments or simply downweights difficult but useful thin-vessel pixels. Table 3 measures the uncertainty-specific gains only on plain UNet++: 0.62 Dice on DCA1 and 1.17 on ARCADE. Without the corresponding full-model ablation and variability estimates, its contribution to the submitted model remains unclear to me. (5) The topology claim is stronger than the evaluation supports. For a vessel segmentation paper, I care about whether difficult branches remain correctly connected, not only whether overall Dice increases. The abstract and contribution list explicitly claim topology evaluation, but the reported metrics are Dice, sensitivity, specificity, HD95 and ASSD. These do not directly establish correct connectivity. The visual examples are helpful, but I do not find them sufficient to support a dataset-level topology claim. (6) The specific frequency assignment is not isolated convincingly. The “w/o FA” variant is not defined well enough to determine whether both branches and comparable capacity are retained. I therefore cannot seperate the benefit of frequency assignment from that of adding a hybrid Mamba2/CNN block. The fixed-kernel argument in Appendix A also does not, by itself, establish the frequency preference of the implemented model. I would put more weight on a matched control than this explanation.

Questions for authors

(1) Which training protocol was actually used for every row of Tables 1 and 3? Please specify source/target image and label access for each method, and reconcile Section 4.2 with Appendix C. This clarification is necessary for me to interpret the main comparison. (2) Are the JSPH-ICA splits patient-disjoint? Which domain supplies validation labels, and are target test images used during adaptation? Please also explain how predictions from the five fold models are aggregated for the reported scores and Wilcoxon tests. (3) What is the main methodological advance beyond combining frequency-aware representation with entropy-weighted prototype adaptation? Please distinguish it from the related ideas in the works above. Which controlled result shows a benefit specific to this coupling, rather than the separate gains of adding the two modules? A convincing answer here would change my assessment of the novelty. (4) What is the ProtoDA versus UPA result on the full FH-Mamba2 model, including fold/run variability? Can you retrospectively examine whether low weights identify incorrect assignments rather than useful thin-vessel pixels, without using target evaluation labels for training or model selection? (5)Can you evaluate the existing predictions with a topology-aware measure such as clDice, together with missed or incorrectly connected branch errors? I would particularly like to know whether the improvements remain in thin-vessel regions, where the paper claims its main benefit. (6) What exactly is removed in “w/o FA”? Please provide a capacity-matched control retaining both operators but removing the frequency split, or swapping their assignments. Under what assumptions is Appendix A intended to explain the implemented Mamba2 block?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.