Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Accurate segmentation of pathological anomalies is often hindered by their high variability and camouflage against anatomically complex backgrounds. While discriminative models struggle with these irregularities, we propose a perspective shift inspired by the radiologist’s cognitive process akin to “Search-by-Subtraction.” We learn a compact subspace from bottleneck background features and use the resulting reconstruction residual energy as an empirical feature-space cue for potential lesions. To leverage this, we introduce the Low-Rank Background Prior (LRBP) Plugin, which isolates anomalies by measuring deviations from a reconstructed background (BG) subspace using a training-time Global Dynamic Memory Bank updated via hard-example mining. Furthermore, to handle varying lesion ambiguity without amplifying noise, we introduce a Discriminability-Guided Gate Supervision mechanism. Here, a lightweight gate predictor learns to adaptively gate residual injection from a ground-truth-derived separability target that estimates foreground-background discriminability during training. Extensive experiments on endoscopic (Kvasir-SEG) and ultrasound (BUSI) datasets demonstrate that our framework achieves strong performance, particularly in high-noise scenarios. Notably, our approach yields substantial reductions in Average Surface Distance, suggesting that modeling background regularity can facilitate precise boundary delineation.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0629_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

Kvasir-SEG (public benchmark used in our experiments): https://datasets.simula.no/kvasir-seg/ BUSI (public benchmark used in our experiments): https://www.kaggle.com/datasets/sabahesaraki/breast-ultrasound-images-dataset ISIC-2018 Task 1 Lesion Segmentation (public benchmark used in our experiments): https://challenge.isic-archive.com/landing/2018/ TN3K (public benchmark used in our experiments): https://github.com/haifangong/TRFE-Net-for-thyroid-nodule-segmentation

BibTex

@InProceedings{CheMin_Injecting_MICCAI2026,
        author = { Chen, Minxin AND Li, Jing AND Zhao, Zian AND Wong, Duo Wai-Chi AND Cheung, James Chung-Wai},
        title = { { Injecting a Low-Rank Background Prior: A Plugin for Medical Image Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16883},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The manuscript proposes a novel, bio-inspired module called the Low-Rank Background Prior (LRBP) Plugin, designed to improve medical image segmentation. Motivated by the clinical “Search-by-Subtraction” cognitive process, the authors model normal anatomical backgrounds as a globally learned low-rank subspace, while pathological lesions are treated as sparse outliers.

  • 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.The translation of the radiological “Figure-Ground Segregation” cognitive process into a computational low-rank modeling problem is highly intuitive and elegant. Moving away from computationally expensive online SVD (common in dynamic MRI/video processing) to a globally learned, dictionary-style memory bank is a smart adaptation for static image segmentation. 2.The LRBP is designed as a true “plugin. “ The authors successfully demonstrate its compatibility with distinct architectural paradigms (ResNet-34, MedNeXt, Swin-Unet). Furthermore, the computational overhead is remarkably low (+0.30% parameters, +0.01% GFLOPs), making it highly practical for clinical deployment.

  • 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.

    1.The motivation of the paper is doubtful. Medical images (like ultrasound and endoscopy) feature highly heterogeneous and noisy backgrounds. Applying a linear PCA to deeply learned, highly non-linear bottleneck features to define a “subspace” is mathematically questionable. 2.The PCA basis B is updated gradient-free every 10 iterations. While computationally efficient, discrete subspace updates during continuous gradient-descent training can sometimes cause training instability or loss spikes. The authors should briefly clarify if any instability was observed and how the periodic update aligns with the momentum of the optimizer. 3.The so-called “Oracle Teacher” is not a neural network; it is simply a hard-coded heuristic formula (Equation 2) that calculates cosine similarity between GT foreground and background. Training a lightweight MLP (Student) to regress this scalar value via Smooth L1 loss is standard supervised regression, not “Knowledge Distillation. “ 4.Table 1 primarily shows improvements over the host backbones themselves. Comparing the method to only two other networks (CMUNeXt, EGE-UNet) is entirely insufficient. The authors must compare their method against state-of-the-art (SOTA) models specifically designed for Kvasir-SEG and BUSI.

    1. Since LRBP is proposed as a plugin, it must be compared against other widely used plug-and-play modules. The comparison with the SE block (Table 1b) is inadequate. How does LRBP compare to spatial attention (e. g. , CBAM), boundary-enhancing modules, or other explicit background suppression/feature recalibration techniques? Without this, the necessity of the complex PCA-memory-bank pipeline is not held.
  • 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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?

    As mentioned in Quenstion 7, the motivation of this paper is doubtful.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    The rebuttal clarifies several conceptual and terminology issues, so my concerns are partially addressed. However, experimental validation, especially against task-specific SOTA methods or stronger plug-and-play baselines still remains unsolved. Therefore, I still find the empirical evidence insufficiently solid to justify raising the score.



Review #2

  • Please describe the contribution of the paper

    This paper proposed a Low-Rank Background Prior (LRBP) Plugin for medical image segmentation, which can be injected into the encoder-decoder segmentation frameworks. Experiments on four datasets have shown the effectiveness.

  • 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.Motivation of this paper seems novel. Rather than directly revise the frameworks, authors try to import new mechanisms named “Search-by-Subtraction” to mimick the radiologist, which appears to have clinical value. 2.The proposed method is plug-and-play mechanism, which could adaptively complete the hard example mining, have shown some generalization ability. 3.Extensive ablation experiments were conducted on the TN3K dataset.

  • 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.

    1.Why is full ablation only performed on the TN3K dataset? Furthermore, the baseline is unclear (is it ResNet-34 U-Net or ResNet-34 Unet-SE?), and apart from the full version, other methods show little improvement or even a decrease in the Dice metric. Please explain why. 2.Why are only the results compared to ResNet-34 U-Net on the ISIC dataset reported? 3.Why are only visualizations of two datasets reported in the Qualitative Results section? 4.The choice of the K value in PCA is only given as an empirical setting, but without a good explanation. 5.Fig. 2 (a) is cited in Section 2.1, but the image and text are too far apart, affecting readability.

  • Please rate the clarity and organization of this paper

    Good

  • 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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?

    The methods presented in this paper are effective, the motivation is clear, and they have certain clinical value. However, there are still some unclear points in the experimental section. If these raised problems can be answered more reasonably, I believe the paper should be accepted.

  • 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.

    The authors have addressed some of my concerns.



Review #3

  • Please describe the contribution of the paper

    This paper proposes a novel plugin named Low-Rank Background Prior (LRBP) for medical image segmentation. The method is motivated by the “search-by-subtraction” paradigm, modeling normal anatomical background as a low-rank subspace and treating lesions as sparse deviations. The approach introduces a global dynamic memory bank to learn background representations via PCA, and extracts residual energy as anomaly cues. Furthermore, a discriminability-guided distillation mechanism is designed to adaptively gate the residual injection. Experiments on Kvasir-SEG and BUSI demonstrate improvements, particularly in boundary quality (ASD).

  • 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)The paper presents a novel perspective by focusing on background modeling rather than foreground discrimination. (2)The proposed LRBP is modular and lightweight, making it easy to integrate into existing architectures. (3)The method shows consistent improvements in ASD, which is clinically meaningful for boundary accuracy. (4)The framework is well-structured, combining subspace learning, residual modeling, and distillation.

  • 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.

    (1)The entire framework is built upon the assumption that background features lie in a low-rank subspace, while lesions are sparse outliers. However, this assumption is not rigorously validated. The provided evidence (PCA cumulative explained variance) is insufficient. High explained variance in PCA is a well-known property of high-dimensional feature spaces and does not necessarily imply intrinsic low-rank structure. (2) The memory bank is constructed exclusively from ground-truth background features, which introduces several critical issues:

    1)The model relies on GT-derived background statistics during training, but such information is unavailable at inference. 2)Any mislabeling in background regions directly contaminates the learned subspace. 3)The memory bank may encode dataset-specific background patterns, limiting generalization. (3) The method reduces the residual signal to a scalar energy value, discarding directional information. This is a significant simplification: Residual vectors may encode richer anomaly cues. The paper claims domain misalignment but does not empirically validate it. This design choice may limit the expressive power of the method. (4) The teacher signal is defined via cosine similarity between FG/BG features, which is heuristic: No theoretical grounding for this choice No comparison with alternative gating strategies Student input is overly simplistic (5) Inconsistent dataset usage in ablation studies. The ablation experiments are conducted on a different dataset (TN3K) than the main benchmarks (Kvasir-SEG and BUSI). This violates the principle of controlled experimentation, as both the model components and the data distribution are changed simultaneously. As a result, it is unclear whether the reported performance differences are attributable to the proposed modules or to dataset-specific characteristics. (6) Lack of statistical significance analysis. The reported improvements, particularly in Dice, are relatively small. However, the paper does not provide any statistical significance analysis (e.g., p-values or confidence intervals) to support that these gains are meaningful rather than due to random variation.

  • 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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?

    Lack of certain theoretical analysis, as well as the lack of experimental analysis.

  • 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



Author Feedback

We thank the reviewers (R1, R2, R3) and meta-reviewer (Meta) for the constructive feedback. We have grouped the major concerns regarding claim calibration, terminology, and experimental framing, and address them below.

Low-rank background prior (R1, R3, Meta). Regarding concerns that applying PCA to non-linear deep features may be mathematically questionable or lacks rigorous validation: we do not claim a theorem that medical backgrounds are globally low-rank in raw image space, nor that lesions are universally sparse outliers. Motivated by search-by-subtraction, LRBP relies on the encoder to model nonlinearity and then uses an explicit linear PCA basis as an efficient, interpretable empirical feature-space prior; kernelized variants can add cost and reduce residual-map interpretability. The operational question is whether scalar residual energy, even without vector direction, is useful for FG/BG separation. The manuscript provides this evidence: a BG-learned basis captures most BG energy with small rank, residual energy separates FG/BG with AUCs of 0.971 (Kvasir-SEG) and 0.890 (BUSI), and median residuals show clear shifts. Thus, our intended claim is an empirical feature-space prior, not a proved intrinsic low-rank structure of medical images.

Memory bank, PCA updates, and inference (R1, R3, Meta). Addressing concerns about reliance on ground-truth (GT) at inference and potential training instability: GT masks are used only during training, specifically for sampling background features, computing the background reconstruction loss, and constructing the GT-guided gate target. At inference, no GT information is used: the memory bank and PCA basis are frozen, and within LRBP, test images are processed by projection onto the learned basis. Furthermore, the PCA basis is a non-trainable buffer updated periodically from the memory bank, not an optimizer-updated parameter. Thus, PCA updates are gradient-free and decoupled from AdamW/backpropagation, mitigating the concern that the basis introduces optimizer-level training instability.

Ablation baseline and TN3K dataset (R2, R3, Meta). Regarding the concern that ablations on TN3K may not be sufficiently controlled or lack a clear baseline: in Table 1(b), the baseline is explicitly the identical ResNet-34 U-Net used in the main study. All LRBP ablation variants use the same TN3K split, host, and training protocol; only the LRBP component changes. The SE block is included only as an additional architectural comparison. TN3K, our largest auxiliary dataset, was used for component analysis to provide a stable setting for rank sweeps and mechanism analysis, while Kvasir-SEG and BUSI were used for the main benchmark evaluation across three host architectures.

Choice of K and empirical thresholds (R2, Meta). Addressing the lack of explanation for PCA K values and heuristics: K=16 was selected empirically and is consistent with Fig. 2(a), where the BG basis reaches 95% cumulative explained variance at around K=10–13 across Kvasir-SEG and BUSI; K=16 is therefore a conservative rank. We also reported a rank sweep in the TN3K ablation (K=4, 16, 64), confirming K=16 offers the best trade-off. The Dice<0.5 criterion is used only for hard-sample mining during training, not as an inference threshold. These are empirical settings, not universal constants.

“Oracle Teacher” terminology and gating (R1, R3, Meta). We agree this is a heuristic, not classical KD. It is GT-guided gate supervision: FG/BG cosine similarity is a separability proxy for when residual injection is reliable. Table 1(b) compares static injection, no teacher guidance, and a heuristic gate. Our intent was privileged supervision, not a separate teacher model.

Overall, these concerns mainly reflect framing issues rather than contradictions in the empirical findings. LRBP remains a decomposition-free inference-time plugin using a learned background basis and adaptive gating to improve boundary accuracy.



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.

    The motivation behind applying low-rank modeling to deeply learned non-linear features should be better justified, along with the assumption that backgrounds lie in a low-rank subspace and lesions are sparse outliers. The authors should clarify the baseline used in the ablation and explain why ablations were performed only on TN3K rather than the main benchmarks. The choice of K in PCA and the empirical thresholds should be better justified. The reliance on GT-derived background features for the memory bank, and how this is handled at inference, should be explained. Finally, the framing of the “Oracle Teacher” as knowledge distillation should be clarified, given that it is a heuristic cosine-similarity formula rather than a learned teacher network.

  • 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.

    The rebuttal addressed the major concerns raised during the original review period. Further comparisons with more baselines and plug-and-play modules can be part of future work.



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.

    The use of the clinically inspired “Search-by-Subtraction” idea is interesting. Employing reconstruction error from an orthonormal basis also appears reasonable. However, the reviewers raise several concerns regarding the methodological details (e.g., the low-rank background assumption) as well as the experimental evaluation. Additionally, the presentation of the paper needs improvement. The authors’ rebuttal partially addresses some of the reviewers’ concerns. Overall, I consider this paper to be borderline acceptable.



Meta-review #3

  • 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.

    Two reviewers Weak Reject (R1=3, R3=3), with R1 maintaining Reject after rebuttal. Core assumption (low-rank background) lacks rigorous validation. Serious experimental design issues including inconsistent ablation datasets.



back to top