List of Papers Browse by Subject Areas Author List
Abstract
Semi-supervised segmentation faces an inherent trade-off between high-level semantic localization and precise boundary delineation due to limited annotated data, which is particularly severe in ultrasound imaging where strong speckle noise and low tissue contrast further degrade representation quality. While Vision Foundation Models (VFMs) such as DINOv3 provide robust global representations, their highly abstracted features tend to suppress fine-grained spatial details, resulting in suboptimal pseudo-label quality in semi-supervised settings. To address this limitation, we propose HG-SemiSeg, a high-frequency–guided feature refinement framework that improves pseudo-label reliability by explicitly enhancing boundary-sensitive information during upsampling. Specifically, we introduce a High-Frequency Component Enhancement (HFCE) module that adaptively selects informative high-frequency components in the frequency domain to provide structural guidance for feature refinement. Built upon a frozen DINOv3 backbone and the feature-agnostic upsampling module AnyUp, HG-SemiSeg leverages the enhanced high-frequency cues produced by HFCE to facilitate spatial detail recovery. This design restores attenuated edge and texture information, thereby improving representational fidelity and pseudo-label quality. Extensive experiments demonstrate that HG-SemiSeg consistently outperforms state-of-the-art semi-supervised segmentation methods. In particular, using only 1% labeled data, HG-SemiSeg achieves average Dice Similarity Coefficient (DSC) improvements of +1.68%, +2.05%, and +1.50% over UniMatch on the CCAUI, BUS-BRA, and PSFHS ultrasound datasets, respectively, while significantly reducing boundary errors. These results validate the effectiveness of high-frequency feature enhancement during upsampling for semi-supervised ultrasound image segmentation. Codes are available at https://github.com/wxmadm/HG-SemiSeg
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3581_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/wxmadm/HG-SemiSeg
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WuXia_HighFrequency_MICCAI2026,
author = { Wu, Xiaming AND Cao, Jiezhang AND Long, Xin AND Rodd, Bardia AND Liang, Yonggang AND Xu, Guoping},
title = { { High-Frequency Guided Feature Refinement for Semi-supervised Ultrasound Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
A high-frequency–guided feature refinement semi-supervised segmentation framework is proposed, that improves pseudo-label reliability by explicitly enhancing boundary-sensitive information during upsampling.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
A high-frequency–guided feature refinement semi-supervised segmentation framework is designed, that improves pseudo-label reliability by explicitly enhancing boundary-sensitive information during up-sampling. Results are relatively good.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
Though results are good, the improvement in fact is not significant. The improvement is less than 2 points averagely.
Moreover, the novelty is quite limited. It seems this work is a combination of existing techniques. Feature frequency analysis has been widely studied in computer vision, image analysis, and image segmentation. It is thus unclear if this work brings enough new contributions to this domain.
Overall, the marginal performance gain and the limited novelty make this paper below the acceptance standard.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Overall, the marginal performance gain and the limited novelty make this paper below the acceptance standard.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper proposed a high-frequency–guided semi-supervised segmentation framework built upon the HFCE module. The proposed approach reinforces boundary-sensitive features to improve structural detail recovery, effectively guides feature upsampling using AnyUp module, and enhances the quality of generated pseudo-labels.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Propose a High-Frequency Component Enhancement (HFCE) module based on a 2D discrete wavelet transform to enhance boundary-aware features and inject high-frequency information into the upsampling stage for improved segmentation accuracy. 2.Adapted a frozen AnyUp upsampling module that leverages both encoder features and HFCE outputs to further refine the upsampling process and preserve structural details. 3.Developed a high-frequency–guided semi-supervised segmentation framework that effectively utilizes unlabeled data to improve model performance. 4.Validated the proposed framework on three diverse datasets.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The paper lacks key methodological details and justification of design choices, including augmentation strategies and feature alignment between modules. The experimental evaluation is limited, with no cross-validation, insufficient analysis of performance under low-label settings, and missing ablation studies to assess individual components. These issues reduce the clarity, reproducibility, and overall rigor of the work.
These are more detailed comments. 1.The manuscript does not specify the augmentation strategies used for the weakly and strongly augmented branches. Clarification on the types and magnitudes of augmentations is necessary for reproducibility and fair comparison. 2.The output of the HFCE module is reported as 3×256×256 and is used as input to the AnyUp module for upsampling. However, this appears inconsistent with the spatial resolution and channel dimensions of the final DINOv3 encoder features. The manuscript lacks an explanation of how this mismatch in feature dimensions is addressed. 3.It is unclear why the AnyUp module is not applied to all upsampling stages within the teacher decoder. The authors should justify this design choice or provide empirical evidence supporting the selected configuration. 4.To better assess the robustness and generalizability of the proposed method, n-fold cross-validation is recommended in all test results. 5.In Table 3, competing methods show substantial performance improvements when using 3% labeled data, whereas the proposed method demonstrates comparatively smaller gains. An analysis explaining this discrepancy would strengthen the paper. 6.An ablation study isolating the contributions of DINOv3 and the HFCE module is missing. Additionally, since AnyUp is introduced to improve upsampling, it would be important to evaluate its effectiveness when applied to all upsampling stages.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I recommend a weak acceptance of this paper. The proposed method presents an pproach to improving segmentation performance by incorporating edge-enhanced, high-frequency features into the decoding process. Despite some concerns regarding experimental details and design choices, the overall contribution is meaningful and has the potential to inspire further research in this direction.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I recommend accepting this paper. The proposed method offers an approach to improving segmentation performance by incorporating a simple and clear edge-enhancement strategy into the decoding process. Although there are some concerns regarding the experimental details and design choices, the overall contribution is meaningful.
Review #3
- Please describe the contribution of the paper
The paper proposes HG-SemiSeg, a semi-supervised segmentation framework for ultrasound imaging. The core claim is that Vision Foundation Models such as DINOv3 produce spatially diffuse features that suppress boundary information, degrading pseudo-label quality in semi-supervised settings. To compensate, the authors introduce a High-Frequency Component Enhancement (HFCE) module that applies a single-level 2D discrete wavelet transform (DWT) to the input image, extracts the three high-frequency sub-bands (LH, HL, HH), applies a stochastic gain per sub-band, concatenates the result, and upsamples to the original resolution. This high-frequency signal replaces the raw image as the guidance input to the AnyUp upsampling module in the weakly augmented branch of a consistency-based semi-supervised pipeline. The framework is evaluated on three publicly available ultrasound datasets (CCAUI, BUS-BRA, PSFHS) at 1% and 3% labeled ratios, consistently outperforming UniMatch and seven other baselines.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Evaluation spans three distinct ultrasound datasets (CCAUI: carotid vessel wall; BUS-BRA: breast tumor; PSFHS: intrapartum fetal anatomy). These targets differ in anatomy, imaging depth, and noise characteristics. Consistent improvements across all three support a claim of generalization within the ultrasound modality.
The motivation is technically grounded. Frozen VFM encoders suppress high-frequency spatial detail in favor of global semantic abstraction. Wavelet sub-bands capture exactly the boundary-related content that frozen encoders attenuate. Replacing raw-image guidance, which carries speckle noise and low tissue contrast, with frequency-filtered cues addresses a real limitation of the AnyUp architecture in noisy imaging settings.
Boundary-sensitive metrics (HD95, ASD) are reported alongside DSC for all datasets and labeling ratios. For anatomical segmentation tasks where contour precision carries clinical meaning, these metrics provide more diagnostic information than DSC alone.
The ablation in Table 4 isolates the contribution of DINOv3, AnyUp, and HFCE as separate components with quantitative results for each combination. This makes the role of each design choice explicit and testable.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The abstract describes HFCE as a module that “adaptively selects informative high-frequency components in the frequency domain.” Equation 2 implements this as an independent scalar gain per sub-band, g_i ~ U(g_min, g_max), sampled without reference to image content. All three sub-bands (LH, HL, HH) are retained unconditionally. Section 2.2 repeats the same framing: “selectively amplifying high-frequency information to provide structural guidance.” The word “selectively” implies content-dependent filtering. The operation is uniform random scaling. This is a factual mismatch between the described mechanism and the implemented one. The abstract and Section 2.2 should be revised to accurately characterize HFCE as stochastic gain augmentation applied uniformly to all wavelet sub-bands.
The uniform distribution U(g_min, g_max) in Equation 2 defines the core HFCE operation. However, hyperparameters g_min and g_max do not appear anywhere else in the paper, making the method not fully reproducible.
Statistical significance is not established. Several key improvements over UniMatch fall within one standard deviation: BUS-BRA 1% reports 78.45±4.36 vs. 75.95±4.74 (difference: 2.5%, SD overlap substantial); PSFHS 1% reports 87.69±0.93 vs. 85.87±1.47.No significance test is reported for any comparison.
Table 4 compares DINOv3 alone, DINOv3 + AnyUp, and the full model (+ HFCE). No condition substitutes an alternative guidance signal for HFCE inside AnyUp (for example, gradient magnitude or Canny edges). The gain assigned to HFCE may arise from providing any structured guidance to AnyUp rather than from the frequency-domain design. An ablation against at least one non-wavelet guidance alternative would isolate this.
PSFHS results (Table 3) are reported without category-wise breakdown. The paper justifies this in the caption by calling aggregated metrics “sufficient.” Fetal head and pubic symphysis differ substantially in size, shape, and imaging contrast. Aggregated DSC and HD95 do not allow the reader to determine which structure drives reported gains, or whether performance degrades on one target while improving on the other.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Three anatomically diverse ultrasound datasets, eight baselines, two label ratios, and boundary-sensitive metrics that align with the claimed mechanism: the evaluation scope is a strength. Results are consistent in direction across all settings.
The primary barrier is the misrepresentation between the abstract and Equation 2.The described operation (adaptive selection of informative components) and the implemented operation (uniform random scaling of all sub-bands) are not the same. Reporting g_min and g_max and adding significance testing are both required for any resubmission. Finally, an ablation against a non-wavelet guidance alternative would substantially strengthen the case for the frequency-specific design.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal addresses several main points. The description-vs-implementation wording issue I raised is conceded; the authors will revise the Abstract and Section 2.2.The PSFHS category-wise numbers are now provided and show consistent gains over UniMatch on both targets. The marginal-Dice critique from R1 is fairly answered by the boundary-metric pivot, where the improvements are substantial and consistent across the three ultrasound datasets.
Two items remain for revision: the non-wavelet guidance ablation I requested was answered with experiments on alternative wavelet bases, which addresses a different question, and no significance testing is provided. Neither undermines the contribution.
I recommend accepting it.
Author Feedback
We thank the reviewers for their insightful feedback. The main concerns and our responses are summarized below: [R1-Q1] Novelty of Paper: Existing semi-supervised segmentation (SSS) methods mainly emphasize semantic consistency while often neglecting fine-grained structural boundaries that are critical for generating high-quality pseudo-labels. Our method is not merely a combination of existing modules; its key novelty lies in demonstrating that high-frequency structural guidance, combined with pre-trained upsampling priors, can significantly improve pseudo-label quality and segmentation performance in low-label ultrasound SSS, which remains insufficiently addressed in current literature. Furthermore, our framework highlights the broader potential of integrating foundation-model priors and generic pre-trained upsampling modules into label-scarce segmentation settings. Overall, our work demonstrates a general framework that effectively leverages pre-trained priors under high-frequency guidance for SSS. [R1-Q2] Limited Performance Gain: In low-label ultrasound SSS, stable boundary recovery is more critical than coarse localization. While Dice gains appear incremental, HFCE significantly improves boundary-sensitive metrics. For instance, on CCAUI (1% labels), HD95 drops from 9.95 to 3.66 and ASD from 3.45 to 1.55.Our method achieves superior stability and boundary precision over UniMatch across all ratios. [R2-Q1] Augmentation Strategies: Weak and strong augmentations strictly follow UniMatch to ensure a fair and direct comparison. [R2-Q2] Feature Mismatch: HFCE outputs serve as high-resolution guidance for the frozen AnyUp module; they are not directly fused with DINOv3 features, thus avoiding explicit alignment issues. [R2-Q3/Q6] Selective Upsampling: AnyUp is applied only at the first upsampling stage to balance performance and efficiency. Deeper integration would significantly increase GPU memory usage and potentially amplify ultrasound speckle noise. [R2-Q4] Lack of Cross-validation: Following standard SSS protocols, we used fixed dataset splits to ensure fair and reproducible comparisons. The consistent improvements observed across multiple datasets and labeling ratios further indicate that the performance gains are not dependent on a specific split. [R2-Q5] Performance Saturation: Baselines often benefit from increased supervision at higher ratios. Our boundary-aware guidance is more effective under extremely sparse supervision, which directly aligns with our motivation for low-label medical imaging. [R2-Q6] Incomplete Ablation: Table 4 presents a progressive ablation from DINOv3 to AnyUp, then to HFCE-guided AnyUp. As HFCE is specifically designed to guide the upsampling framework, it cannot operate independently. We provided more detailed analysis on our Github page. [R3-Q1] Misleading Description: We appreciate the observation and will revise the Abstract and Section 2.2 for clarity. [R3-Q2] Missing Hyperparameters: The parameters g_min and g_max are fixed across all experiments. Optimal performance is consistently observed at g_min and g_max=5.[R3-Q3] Lack of Significance Testing: Our method consistently achieves lower HD95/ASD, indicating superior structural stability. The larger standard deviations in BUS-BRA (1% setting) are inherent to the dataset’s blurry boundaries and strong speckle noise under sparse labels. [R3-Q4] Insufficient Baseline Comparison: Additional experiments with different wavelet bases (e.g., db3 and coif3) show that our method consistently outperforms competing approaches across settings. Moreover, the compared methods were selected as representative SOTA methods widely used in recent literature. [R3-Q5] Missing Class-wise Evaluation: PSFHS category-wise results (omitted for space) show consistent gains in the 1% setting: our method leads UniMatch on both Fetal Head (84.41% vs. 82.74%) and Pubic Symphysis (90.96% vs. 88.99%). Additional details, results, and code will be released on GitHub.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
Three reviews note marginal gains and limited novelty, but methodological concerns dominate.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
While the integration of high-frequency features into semi-supervised segmentation is conceptually well-motivated, consensus post-rebuttal highlights that the manuscript’s core novelty remains limited, representing an incremental combination of pre-existing modules. Furthermore, critical concerns remain unresolved regarding the factual mismatch between the described adaptive mechanism and its actual stochastic formulation, the lack of ablation comparisons against non-wavelet spatial guidance baselines, and marginal empirical gains that lack statistical significance testing.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After the rebuttal, most reviewers provided positive recommendations for this work. Overall, although considering high-frequency information into semi-supervised segmentation is not new, it is reasonable and effective. By using DINO features and wavelet operations, this work achieves superior performance, particularly on shape-related metrics. It indicates a clear contribution and could benefit the relevant community. Therefore, acceptance is recommended, and the authors are suggested to improve the manuscript based on reviewers’ comments in the camera-ready version.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper presents a semi-supervised ultrasound image segmentation framework enhanced by high-frequency component guidance. While the method is relatively straightforward, which is composed of a few established modules (e.g., the teacher-student framework, the discrete wavelet transforms, and the AnyUp upsampling module), the design motivation is sound. I am also impressed by the comprehensive evaluation settings, which include three distinct ultrasound image datasets and eight competing baselines under two label ratios (although it sacrificed some clarity regarding the methodology and implementation details). After the rebuttal, the author generally addressed the primary concerns raised by the reviewers. The unaddressed comments do not undermine the core empirical strength of the work, and thus I recommend accept. PS: I have one additional comment for the authors’ potential future revisions: the high-frequency components in the ultrasound may not fully correspond to the object boundaries as there is severe noise existing in the ultrasound image. Emphasizing the high-frequency component in the ultrasound data may also amplify the noisy level.