acceptodds
Under review as a conference paper at ICLR 2027

FReMCI: Frequency-Reliable Memory Cross Intervention for Semi-Supervised Ultrasound Image Segmentation

Abstract

Semi-supervised ultrasound image segmentation is clinically attractive because dense annotations are costly, but probe-dependent appearance, speckle noise, low contrast, and ambiguous boundaries make unlabeled supervision unreliable. Existing methods can propagate errors through pseudo labels, spatially mixed regions, and weak foreground features. We present FReMCI, a frequency-reliable memory cross intervention framework that controls cross-sample transfer at frequency, region, and feature-memory levels. Phase-Preserved Low-Frequency Amplitude Sharing (PLAS) extends Fourier amplitude interpolation to paired labeled–unlabeled images while retaining each image's phase and high-frequency amplitude. Reliability-Guided Semantic Region Intervention (ReSRI) combines confidence, entropy, and class margin from an EMA teacher with teacher–student agreement to construct reliability-weighted pseudo targets and foreground-centered bidirectional region exchanges with seam-suppressed supervision. Memory Cross Injection (MCI) stores labeled and reliable unlabeled foreground features and injects opposite-source context into the student decoder; teacher predictions remain independent of the memory state. Inference uses the single student segmentation network. Under the reported fixed, single-run protocol on five ultrasound tasks from four dataset sources, FReMCI ranks highest in average Dice across all three labeled-data protocols (1/10, 3/10, and 1/2), outperforming the strongest competing averages by up to 3.87 points.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Ultrasound imaging, Medical image segmentation, Semi-supervised learning, Uncertainty estimation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_5RUKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes FReMCI, which combines low-frequency amplitude sharing (PLAS), reliability-guided region mixing (ReSRI), and foreground memory injection (MCI) for semi-supervised ultrasound segmentation. The authors report the highest average Dice on five tasks from four dataset sources at three labeling ratios with eighteen baselines. They also provide component ablations and parameter sensitivity analysis.

Critical strengths
  1. Promising results on semi-supervised segmentation. Tab. 1 to Tab. 3 report the highest average Dice at different labeling ratios. Tab. 4 also shows gains from different module combinations.
  2. Simple inference. The additional modules are used only during training and do not add steps at inference.
Critical weaknesses
  1. The experiments do not show reliable advantages over the baselines. Actual baseline settings and training costs are not fully reported, making it difficult to assess the trade-off between performance and training cost across methods. The experimental discussion mainly repeats table scores without analyzing trends or outliers. In particular, it does not explain why Dice on UDIAT drops when the labeled fraction increases from 3/10 to 1/2. Missing setup details, such as data splits, and the encrypted supplement make the codes difficult to check and reduce confidence in reproducibility. These gaps leave it unclear whether the gains are stable and worth the extra training cost.

  2. The method description and ablations do not fully support the main design claims. The motivation sounds reasonable, but the specific claims in the Introduction need clearer support from prior work or direct experiments. Appendix D claims independence from memory, but Eq. 7 and Alg. 1 use a student prediction affected by memory to compute reliability. The figures also contain conflicting or missing connections. Fig. 1 does not clearly label the PLAS and ReSRI modules or show which operations belong to each module. In Fig. 2, "Top-k extract" is connected directly to the input image, while Section 3.3 states that the top-1 and top-2 values come from the teacher probabilities. Table 4 shows gains from some module combinations, but all variants share the reliability estimator and PLAS is only tested with ReSRI. These tests do not fully explain why the specific design choices are needed or whether the modules address the problems described.

Questions for authors
  1. Supplement and data splits. Please provide accessible supplementary code and explain the construction of the Aneil/Aryan subsets, data splits, and labeled-sample selection. Are patients or image sequences kept separate across splits where relevant? Are test sets fixed across labeling ratios, and is each smaller labeled set included in the larger one?

  2. Baseline results and training cost. Were the baseline results rerun for this study or taken from previous papers? If rerun, did all methods use the same data splits, labeled samples, preprocessing, and evaluation procedure? If taken from previous papers, please identify those results and explain whether their settings match this study. Please also report the actual backbones, pretraining, training budgets, runtime, and peak memory for FReMCI and the strongest baselines. Are the gains still present under comparable training budgets?

  3. Results across labeling ratios. The average Dice advantage over the strongest baseline becomes smaller as the labeled fraction increases. Is this trend consistent across repeated runs? Please provide a comparison with a strong baseline at a higher labeling ratio, such as 3/4, and a 1/1-label fully supervised reference under comparable settings. How does the trade-off between labeling effort, training cost, and segmentation performance change across these ratios? Please also explain the UDIAT performance drop from 3/10 to 1/2 labels.

  4. Method details and evidence. Appendix D states that pseudo-label computation is independent of memory, but Eq. 7 and Alg. 1 use a student prediction affected by memory to compute reliability. Please explain this difference and make the text, figures, and code consistent. What studies or experiments support the problems described in the Introduction and show that the modules solve them?

1 comment

I thank the authors for the response and additional experiments.

However, some concerns are still not resolved. Splits are image-level without patient or sequence separation. Training cost and budget-matched comparisons cover one baseline on two datasets, and their gains lie within the reported deviations. Table 4 is still a single run, and the new reliability controls differ by less than one deviation. The protocol details appear only in the response and not in the revised paper.

In addition, the supplement does not seem to contain code for the new experiments.

In conclusion, I maintain my rating.

Reviewer_j68FOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The FReMCI proposed in this paper combines low-frequency amplitude sharing, region mixing based on reliability guidance, and foreground memory mechanism for semi-supervised ultrasound image segmentation. In experiments covering four sources, five tasks, and three label ratios, this method performed better than other comparison methods in a single run, achieving the highest average Dice and Jaccard coefficients.

Critical strengths

The results of the paper show significant performance improvement, especially in the case of scarce annotated data (one tenth), which is a good inspiration for utilizing insufficient data for training.

Critical weaknesses
  1. Although the experimental section provides comparative and ablation experiments, it does not fully explain how these results support key design choices and core claims. Therefore, the performance improvement presented lacks sufficient strong evidence to support it.

  2. The introduction section fails to fully point out the existing research gaps and does not clarify what significant progress the proposed framework has made. Based on the current presentation, both the research motivation and novelty lack sufficient motivation.

Questions for authors

If the teacher and student agree on an incorrect prediction, could the shared reliability score propagate that error through pseudo label? How does the framework control this risk?

Reviewer_ZCh4OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper proposes FReMCI, a semi-supervised ultrasound segmentation framework combining frequency-domain augmentation, reliability-weighted pseudo-labeling, and cross-sample feature memory. The method demonstrates promising improvements across five segmentation tasks, particularly when labeled data are scarce.

However, the main limitation is the absence of supervised-only learning curves and a fully supervised baseline. Without these comparisons, it is difficult to quantify the actual benefit of unlabeled data or determine how closely semi-supervised learning approaches fully supervised performance.

Furthermore, the method assumes that manually annotated labels are reliable, despite the inherent ambiguity of ultrasound boundaries. Evaluating robustness to annotation noise and inter-observer variability would strengthen the work.

Overall, the empirical results are encouraging, but the limited methodological novelty and incomplete experimental validation weaken the contribution.

Critical strengths

Strong empirical performance: The method achieves the highest average Dice scores across five ultrasound segmentation tasks and three labeling ratios, with particularly encouraging improvements under limited supervision.

Complementary design: The integration of frequency-domain augmentation, reliability-guided pseudo-labeling, and feature memory addresses practical challenges in ultrasound segmentation.

Extensive baseline comparisons: Evaluation against 18 semi-supervised methods provides useful empirical evidence, although the absence of fully supervised baselines remains a limitation.

Efficient inference: All additional training modules are removed at inference, leaving a standard segmentation network.

Critical weaknesses

Missing supervised baselines: The absence of supervised-only learning curves and a fully supervised (100% labeled) baseline makes it difficult to quantify the actual benefit of unlabeled data and the method's annotation efficiency.

Assumption of accurate annotations: The framework treats manual labels as fully reliable, despite the ambiguous boundaries and inter-observer variability inherent in ultrasound imaging. Robustness to annotation noise is not investigated.

Limited methodological novelty: The methodological novelty appears somewhat limited, as the proposed framework builds upon established techniques, including Fourier-based augmentation (Yang & Soatto, 2020), confidence-guided pseudo-labeling (Sohn et al., 2020), bidirectional copy-paste (Bai et al., 2023), and cross-image prototype learning (Ding et al., 2025). While their integration may be useful, the paper would benefit from a clearer demonstration of what is fundamentally new beyond these existing approaches.

Insufficient statistical validation: Results are reported from single runs without variance estimates or statistical significance testing.

Incomplete ablations: The experiments do not isolate several important design choices, including individual reliability terms and memory-injection mechanisms.

References

Yang, Y., & Soatto, S. (2020). FDA: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 4085–4095).

Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C. A., Cubuk, E. D., Kurakin, A., & Li, C.-L. (2020). FixMatch: Simplifying semi-supervised learning with consistency and confidence. In Advances in Neural Information Processing Systems, 33, 596–608.

Bai, Y., Chen, D., Li, Q., Shen, W., & Wang, Y. (2023). Bidirectional copy-paste for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 11514–11524).

Ding, Z., Chen, G., Zhang, Q., Wu, H., & Qin, J. (2025). CSC-PA: Cross-image semantic correlation via prototype attentions for single-network semi-supervised breast tumor segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 15632–15641).

Questions for authors

Fully supervised comparison: Could the authors provide supervised-only results at 10%, 30%, and 50% labeled data, along with a fully supervised baseline using 100% of the available training annotations? This would clarify the actual benefit of unlabeled data and the annotation efficiency achieved.

Annotation reliability: How robust is FReMCI to noisy or inaccurate manual annotations? Have the authors considered inter-observer variability, particularly around ambiguous ultrasound boundaries?

Methodological novelty: Given the existing literature on Fourier augmentation, confidence-based pseudo-labeling, bidirectional copy-paste, and cross-image feature learning, what is the principal methodological innovation beyond combining these established techniques?

Reliability estimation: What is the justification for multiplying confidence, entropy, prediction margin, and teacher–student agreement? Could the authors compare this formulation with simpler confidence-based filtering?

Statistical significance: Could the authors report results across multiple random seeds, including standard deviations or confidence intervals, to establish the reliability of the observed improvements?

Computational efficiency: What additional training time and memory overhead does FReMCI introduce compared with the strongest baselines?

Comments on the paper

1 comment

Dear reviewers,

The authors have submitted their rebuttal. Please review their responses and provide any additional comments or updates. I would appreciate your engagement with the discussion.

Thanks, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.