List of Papers Browse by Subject Areas Author List
Abstract
Brain tumor segmentation (BraTS) of 3D Magnetic Resonance Imaging (MRI) aims to facilitate clinical analysis of brain cancer. Existing BraTS segmentation works tend to exploit convolutional neural networks (CNNs) or vision transformers (ViTs), yet CNNs have a restricted receptive field that focuses on local context only and ViTs suffer from high computational overheads due to quadratic complexity. Recently, Mamba has shown superior performance over ViTs in long-range dependency modeling, offering linear computational complexity and lower memory consumption. However, these methods primarily learn feature representation in the spatial domain, overlooking valuable heuristics embedded in the frequency domain. Inspired by this, we propose BraTS-UMamba, a novel Mamba-based U-Net designed to enhance brain tumor segmentation by capturing and adaptively fusing bi-granularity based long-range dependencies in the spatial domain while integrating both low- and high-band spectrum clues from the frequency domain to refine spatial feature representation. We further enhance segmentation through an auxiliary brain tumor classification loss. Extensive experiments on two public benchmark datasets demonstrate the superiority of our BraTS-UMamba over state-of-the-art methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0487_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)
N/A
BibTex
@InProceedings{YaoHao_BraTSUMamba_MICCAI2025,
author = { Yao, Haoran and Xiong, Hao and Liu, Dong and Shen, Hualei and Berkovsky, Shlomo},
title = { { BraTS-UMamba: Adaptive Mamba UNet with Dual-Band Frequency based Feature Enhancement for Brain Tumor Segmentation } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15975},
month = {September},
page = {97 -- 106}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors proposed BraTS-UMamba, a novel Mamba-based U-Net designed to enhance brain tumor segmentation by capturing and adaptively fusing bi-granularity based long-range dependencies in the spatial domain while integrating both low- and high-band spectrum clues from the frequency domain to refine spatial feature representation. They further enhanced segmentation through an auxiliary brain tumor classification loss.
- 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.
The authors introduced a few new designs in the deep learning model, and conducted ablation studies to show the effectiveness of these designs.
- 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.
-
In the results section, they didn’t compare with nnUNET, which has proven to be adaptive to a large varieties of dataset, and can generate good results without requiring to tune the model.
-
In Figure 3, it seems that the authors nick-picked only large lesions and those ones who has a 0.95 Dice for both examples. It would be nice if the authors can select random results, and any lesion size.
-
- 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.
- 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/2025/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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The authors propose new designs of deep learning network, and they prove to work.
- 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
Authors proposed the first study using new Mamba approach for brain tumor segmentation and compare the performance of this approach to previous models. To realise this study the authors used MSD and BRATS public database. The purpose of the paper is interesting to clinical practice. Collaboration with clinician could really improve the interest of the study (with clinical evaluation, clinical workflow gain…)
- 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.
The authors proposed innovative approach using Mamba for brain tumour segmentation which was not tested before
- 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.
- no clinical validation of the segmentation was performed. It could be very interesting to ask radiotherapist or other medical doctor to evaluate the quality of the segmentation rather than metric as DICE which are questionable and can be far away from the clinical requirement.
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
- As mentioned before, DICE index is very questionable and couldn’t be use as unique metric to evaluate the performance of the model. Other metric should be use or at least, this point should be discussed in the discussion section.
- As a perspective, clinical evaluation of the segmentation by a radiotherapist should be mentioned.
- The proposed model should be tested in a clinical centre with clinical routine rather than public well-structured data.
- Doing that will allow the authors to evaluate the interest of the proposed method in term of working time gain using that method in comparison to manual segmentation done by the radiotherapist.
- 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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The authors proposed new Mamba approach for brain tumour segmentation which is, to my knowledge and author’s one, the first study using this approach. However, only simple metrics as DICE was used to evaluate the performance of the model which is not enough to validate the approach.
- Reviewer confidence
Very confident (4)
- [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 #3
- Please describe the contribution of the paper
The paper presents BraTS-UMamba, a novel Mamba-based U-Net architecture specifically designed for brain tumor segmentation in 3D MRI images. The main contributions can be summarized as follows: Novel Architecture with Adaptive Mamba Module The paper introduces a U-shaped encoder-decoder architecture that leverages the Mamba state space model to overcome limitations of both CNNs (restricted receptive field) and Vision Transformers (high computational complexity). The core innovation is the Adaptive Mamba (AdM) module, which captures bi-granularity based long-range dependencies with different perspectives and uses a soft, adaptive fusion approach rather than hard fusion used in previous Mamba methods. Dual-Domain Feature Enhancement The paper proposes a unique approach that combines spatial and frequency domain information. While most existing methods focus solely on spatial features, BraTS-UMamba introduces the Frequency Guidance based Feature Enhancement (FGFE) module that leverages both low-band spectrum (representing global structures) and high-band spectrum (representing edges and textures) to refine spatial feature representation. This dual-domain approach helps better capture tumor boundaries and complete regions. Multi-Scale Feature Extraction To address the challenge of capturing features from tumors of various sizes, the model incorporates multi-scale convolutions within the AdM module, allowing it to extract local details at different scales. This approach is particularly effective for small tumors with blurred boundaries, which are common in brain MRI data. Auxiliary Classification Approach The paper implements an auxiliary brain tumor classification loss in addition to the primary segmentation loss. This helps the network focus on the tumor regions despite their typically small size compared to normal brain tissue, addressing the class imbalance problem inherent in brain tumor segmentation.
- 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.
Strengths of the Adaptive Mamba (AdM) Module:
Novel Bi-Granularity Feature Extraction with Adaptive Fusion: AdM introduces a unique approach to capturing long-range dependencies through bi-granularity based global features Unlike other Mamba methods using hard fusion, AdM employs adaptive fusion through attention mechanisms, allowing dynamic adjustment of feature importance
Multi-Scale Convolution Integration for Local Detail Capture: Incorporates parallel convolutions with various kernel sizes (1×1×1 to 7×7×7) to extract multi-scale local features This approach effectively addresses the challenge of small tumors with blurred boundaries in brain MRI images
Strengths of the Frequency Guidance based Feature Enhancement (FGFE) Module:
Cross-Domain Integration of Frequency Information: FGFE represents a significant innovation by incorporating frequency domain information to enhance spatial features While most existing methods operate solely in the spatial domain, this dual-domain approach captures complementary information that improves segmentation quality
Specialized Dual-Band Spectrum Utilization: Leverages both low-band spectrum (for global structures and continuous regions) and high-band spectrum (for edges and textures) This dual-band approach precisely addresses the needs of brain tumor segmentation, where both complete tumor regions and accurate boundaries are critical
- 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.
There’s a fundamental inconsistency in how the manuscript claims to solve the problem of small tumor detection versus how the model is actually evaluated. The authors repeatedly emphasize that a key advantage of their approach is the ability to better capture small tumors with blurred boundaries through:
Multi-scale convolutions in the AdM module to extract local details at various scales Frequency domain information in the FGFE module to enhance boundary representation Auxiliary tumor classification loss to address class imbalance
However, the evaluation metrics used (DSC and HD95) are aggregate measures that heavily favor performance on larger tumors rather than small ones. This creates a disconnect between the claimed contribution and the evidence provided:
The Dice Similarity Coefficient is well-known to be biased toward larger structures, as small segmentation errors in large tumors have minimal impact on the overall score, while small tumors are disproportionately penalized by even minor errors. The evaluation doesn’t include any metrics specifically designed to assess performance on small tumors, such as sensitivity analysis on tumors below a certain size threshold or precision-recall curves for small lesions. Despite claiming the model addresses the challenge of small tumors, there are no visualization examples specifically showing successful segmentation of small tumors compared to baseline methods.
This mismatch between the claimed contribution (better performance on challenging small tumors) and the evaluation methodology (metrics that primarily reflect performance on larger tumors) represents a significant logical flaw in the manuscript. Without evidence specifically targeting small tumor performance, the claimed advantage of the multi-scale approach in the AdM module and the boundary enhancement in the FGFE module remains unsubstantiated.
- 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.
- 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/2025/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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
- Innovative Technical Contribution The paper presents two novel modules (AdM and FGFE) that meaningfully extend Mamba’s capabilities for brain tumor segmentation. The adaptive fusion approach and dual-domain feature enhancement represent genuine advances beyond simply applying existing techniques.
- Strong Empirical Results Testing on two public datasets shows significant performance improvements over seven competitive baselines, with consistent gains in both DSC and HD95 metrics. The comprehensive comparisons and ablation studies provide solid evidence for the method’s effectiveness.
- Clear Presentation and Potential Impact The paper is well-structured with informative figures and clear explanations. The approach addresses real clinical challenges in tumor segmentation while maintaining reasonable computational complexity, making it potentially valuable for practical applications.
- Reviewer confidence
Very confident (4)
- [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 and meta-reviewer for their thoughtful comments and helpful suggestions. Below, we respond to each point for clarification and transparency.
Comparison with nnUNet We appreciate the suggestion to include nnUNet as a baseline. While nnUNet is indeed a powerful and adaptive method, we focused our comparisons on methods that are more structurally aligned with our proposed network, to better highlight the contributions of specific components. Nevertheless, we agree that including nnUNet would provide valuable additional context, and we will consider this comparison in future work to further validate our method.
Visual results in Figure 3 (selection bias) We appreciate the reviewer’s concern regarding potential cherry-picking. Our intention was to showcase qualitative examples that clearly demonstrate the behavior of different methods with respect to boundary quality and segmentation mask connectivity. While larger lesions are more visually discernible, we agree that including smaller or randomly selected cases would improve representativeness. We appreciate this valuable suggestion and will adopt a more diverse sampling strategy in future experiments.
Evaluation metrics vs. small tumor claims We appreciate the reviewer’s detailed analysis and insightful feedback. Our work aims to address challenges in small tumor segmentation through architectural innovations (adaptive Mamba and frequency-guided feature enhancement). While DSC and HD95 are standard metrics, we acknowledge that they are biased toward larger structures. Due to space constraints, we were not able to include stratified evaluations specifically on small tumors. Nonetheless, we believe that our model’s design remains well justified in targeting small-lesion challenges, and we view the incorporation of size-specific metrics and detailed analysis as important directions for future work.
Clinical validation We agree that clinical validation is a critical step toward practical deployment. Our current study focuses on algorithmic development and quantitative evaluation, which are necessary first steps. We fully acknowledge that clinical expert assessment can offer complementary perspectives, and we aim to incorporate expert feedback in future collaborations to better assess clinical relevance and alignment.
We again thank the reviewers for their constructive feedback. We hope this response clarifies our design choices and evaluation strategy.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
Due to many raised advantanges, all three reviewers are positive to accept this paper, and their ratings are ‘weak accept’, ‘weak accept’, and ‘accept’. Hence, this work can be accepted now. Meanwhile, the authors are suggested to revise the paper based on these reviewer comments.