List of Papers Browse by Subject Areas Author List
Abstract
In recent medical image analysis, Convolutional Neural Networks (CNNs) and State Space Models (SSMs) have set benchmarks in segmentation tasks. While CNNs excel in capturing local fine-grained features, SSMs achieve remarkable global context understanding with linear complexity. However, Mamba-UNet, a pioneering pure SSM-based model, still exhibits deficiencies in feature representation, fusion efficiency, and spatial detail reconstruction. To address these limitations, we propose VM-NeXT UNet, an improved architecture that synergizes ConvNeXT with VSS Blocks.VM-NeXT UNet introduces four key optimizations: (1) a dual-encoder parallel structure for comprehensive feature extraction; (2) a channel-spatial attention gating module in skip connections for adaptive feature screening; (3) multi-scale convolution layers in the decoder to preserve spatial details; and (4) a combined FocalLoss and DiceLoss strategy to focus on hard samples. Experiments on the Synapse and ACDC datasets yielded Dice scores of 86.21% and 92.42%, respectively. The results demonstrate that VM-NeXT UNet significantly outperforms the original Mamba-UNet and achieves competitive performance against state-of-the-art methods, highlighting its potential for reliable clinical deployment.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4748_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{CaoBox_VMNeXT_MICCAI2026,
author = { Cao, Boxuan AND Jiang, Peilin},
title = { { VM-NeXT UNet: Synergizing ConvNeXT and Visual Mamba for Robust Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes VM-NeXT UNet, a hybrid U-shaped architecture that combines ConvNeXT blocks for local feature extraction with Visual State Space (VSS) blocks for global context modeling. The architecture uses a dual-encoder parallel design with feature injection blocks for inter-branch communication, a Channel-Spatial Attention Gating (CSAG) module for skip connections, a multi-scale decoder, and a combined FocalLoss + DiceLoss training objective. Experiments are conducted on Synapse (multi-organ CT) and ACDC (cardiac MRI) datasets.
- 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 dual-encoder parallel design is a reasonable approach to combining the strengths of CNNs (local features) and SSMs (global context), with the lightweight injection blocks providing a computationally efficient fusion mechanism. -Strong quantitative results on Synapse (DSC 86.21%, HD95 16.68) outperforming all listed baselines including Cascaded MERIT. -The CSAG module for skip connections is well-designed, addressing the known semantic gap problem in encoder-decoder architectures. -The ablation study (Table 3) systematically demonstrates the contribution of each component.
- 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.
-Limited novelty: each individual component (dual-encoder, attention gating, multi-scale convolutions, Focal+Dice loss) is well-established. The contribution is primarily architectural assembly rather than methodological innovation. -The claim of “Title Suppressed Due to Excessive Length” appearing in the running header suggests formatting issues that indicate rushed preparation. -No efficiency analysis: the paper never reports parameter count, FLOPs, or inference time. Given that the dual-encoder design doubles the encoder computation, this omission is significant. -The comparison is limited to segmentation-focused baselines. No comparison with other hybrid CNN-Mamba architectures or recent medical segmentation methods beyond those listed.
- 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 provide sufficient information for 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?
I lean toward rejection because, while the paper is technically solid and shows strong empirical performance, its overall contribution is limited in terms of novelty and completeness. The proposed architecture mainly combines well-established components (dual encoders, attention gating, multi-scale decoding, and standard loss functions) without introducing a clearly new methodological insight. Although the results on Synapse and ACDC are competitive, they are demonstrated on only two saturated benchmarks, which makes it difficult to assess generalization or practical impact.
In addition, important aspects of the evaluation are missing, particularly efficiency metrics such as parameter count, FLOPs, and inference time, this is especially critical given the computational cost implied by the dual-encoder design
- 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 provides useful clarifications but does not change the fundamental assessment. The novelty remains incremental, the efficiency argument is actually weakened by the numbers provided.
Review #2
- Please describe the contribution of the paper
The paper presents a DNN architecture for 2D multi-class semantic segmentation in medical images, combining CNN-based and Mamba-based components, and evaluates it on the Synapse and ACDC datasets, where it reports performance gains over existing methods.
- 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 method achieves performance gains over strong CNN, transformer, and SSM baselines across Synapse and ACDC, indicating competitive empirical effectiveness despite the limited evaluation scope.
The paper attempts to tackle important challenges in medical image segmentation, such as preserving fine structural details and improving robustness across complex anatomical patterns.
- 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 is not written clearly enough; some parts feel repetitive, especially the overlap between the method description and Section 2.1 on the overall architecture, which makes the presentation less concise.
Several motivation claims are too strong relative to the evidence provided; for example, the paper states that patch-wise tokenization in ViT loses fine-grained spatial information and leads to blurred boundaries for small anatomical structures, but it does not provide direct experiments, targeted analysis, or citations to support this claim. The discussion of prior work is also not sufficiently supported; the paper claims that Mamba-UNet has four inherent limitations, but these points are not convincingly validated through evidence, citations, analysis, or ablation.
There is also a mismatch between the motivation and the evaluation setup; the paper argues that SSM-based models are suitable for high-resolution medical images, yet the experiments appear to use 224×224 inputs, so this motivation is not well justified by the reported setting.
The loss design is also not clearly validated; although the paper motivates the use of Dice Loss + Focal Loss , the ablation does not clearly specify which loss is used when this component is removed, and it would be more convincing to include direct comparisons with BCE, Dice, and Focal losses individually, so that improvements from the architecture can be separated from those due to loss optimization.
I am also curious about the specific contribution of the injection block; the paper does not provide enough targeted ablation to show whether the reported gains actually come from this component, and a direct comparison with and without the injection block would make its role much clearer.
The qualitative analysis is somewhat incomplete; it does not discuss convolution-based baselines, even though local detail preservation is one of the main motivations, so including CNN-based visual comparisons would better support the claimed visual advantages of the method. The qualitative evaluation is further limited because no qualitative results are shown for the ACDC dataset; adding visual examples from ACDC would help verify whether the claimed advantages hold consistently across both evaluated datasets.
The paper also lacks statistical significance analysis; it reports only point estimates and does not provide standard deviation, confidence intervals, or results from multiple runs, making it difficult to assess whether the improvements are stable and reliable rather than partly due to random variation.
The paper does not report efficiency metrics such as parameter count or GFLOPs; without these results, it is difficult to assess the practical cost of the proposed design or determine whether the performance gains come with a reasonable computational tradeoff.
- Please rate the clarity and organization of this paper
Poor
- 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?
My concerns are mainly about novelty and writing quality. Although the paper reports performance gains, many key claims are not supported by targeted experiments, citations, or ablation. Important components, such as the loss design, injection block, and other modules, lack clear ablation, making their contributions uncertain. Additionally, there is no statistical significance analysis.
- Reviewer confidence
Very confident (4)
- [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.
Although some concerns remain only partially resolved, I believe they are manageable and could be further strengthened in the final version. In particular, while combining existing CNN, attention, and Mamba-based components is a common and reasonable direction in medical image segmentation, especially for UNet-style architectures aiming for incremental performance improvements, such approaches would benefit from more carefully isolated ablation studies to clearly validate the contribution of each integrated module. The rebuttal still does not fully disentangle architecture gains from optimization gains, since targeted and isolated ablations for the proposed loss strategy and the injection module remain limited. This is particularly relevant because relatively small performance improvements in segmentation tasks can sometimes arise primarily from loss-function optimization alone. Additional isolated comparisons between BCE, Dice, Focal, and combined loss formulations would help clarify how much improvement originates from optimization choices versus architectural design. Statistical significance analysis across multiple runs would also improve confidence in the robustness and consistency of the reported gains. Furthermore, the added efficiency analysis shows that the proposed model introduces additional parameters, FLOPs, and latency compared with Mamba-UNet, although the observed accuracy improvements may justify this moderate tradeoff.
Despite these remaining limitations, after considering the rebuttal, I lean weakly toward acceptance. The authors addressed several concerns by improving the clarity of the paper, adding efficiency metrics, and further explaining the motivation behind the proposed architecture. The reported performance on Synapse and ACDC remains competitive relative to strong CNN, transformer, and Mamba-based baselines, and the overall framework appears technically reasonable. Additionally, the clarifications and analyses provided in the rebuttal improve the completeness of the work, and these details must be incorporated into the final version of the paper. Overall, I believe the paper reaches the acceptance threshold, although incorporating the additional analyses, ablations, and clarifications discussed in the rebuttal would make the work more convincing.
Review #3
- Please describe the contribution of the paper
Authors propose a hybrid deep learning architecture that addresses limitations of Mamba‑based and state‑space models by incorporating components such as Dual‑Encoder Parallel Hybridization, Adaptive Feature Gating, Multi‑Scale Reconstruction, and an Optimized Loss Strategy.
- 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 paper is well written and structured, and the methodological gap in current state‑of‑the‑art methods is clearly articulated. The proposed approach combines the components described above to address this gap, introducing a novel hybrid architecture.
- 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 architecture is relatively complex, which is manageable for 2D segmentation, but extending it to full 3D applications may introduce substantial computational overhead that the authors should examine in future work.
- 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
- The Synapse dataset includes more anatomical labels, and the authors should clarify in the Dataset Description section why only eight classes were selected for the experiments.
- There should be visual comparison details on figure 1 highlighting how the methods differ across specific structures—for example, comparing the right kidney results of method X with those of method Y—to guide the reader in observing these differences.
-
It is understandable that there is limited space to show results for all methods listed in Table 1.However, it would be more consistent to present qualitative comparisons for the top three SOTA methods ScaleFormer, Parallel MERIT, and Cascaded MERIT. The figure should be updated accordingly to match the methods highlighted in the table.
-
Figure 1 uses fonts that are too small to read clearly. The authors should increase the font size to improve visibility and overall readability.
- The ablation study should be presented before the comparison with SOTA methods, as it is part of the model design process and improves the logical flow of the manuscript.
- The GPU RAM used should also be mentioned here, as it provides a clearer sense of the computational burden.
- All components used in the method should be cited, including VSSM, MIRB, and others.
- 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?
The authors clearly identify the methodological gaps and make effective use of the helpful deep architectural components.
- Reviewer confidence
Very confident (4)
- [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 acceptance, as the proposed architecture offers a clear methodological contribution by addressing the lack of local‑detail modeling in pure Mamba designs through a well‑integrated ConvNeXT-Mamba synergy, resulting in a hybrid model that is both practically effective and conceptually meaningful.
Author Feedback
Thank the AC and reviewers.Our point-by-point responses follow.
1.Clarification of Novelty & Addressing Unsupported Claims (R1, R2, Meta) Regarding the concern that our work is merely simple architectural assembly: We clarify our work is far from trivial assembly.Our core innovations include three aspects: first,synergistic ConvNeXT-Mamba dual-encoder fusion,where Injection Block and bottleneck cross-attention enable inter-branch interaction, preventing heterogeneous feature misalignment; second,we devise a multi-scale enhanced decoder with our Multi-Scale VSS Block,addressing single-scale limitations and improving multi-scale anatomical segmentation;third, we adopt Focal Loss replacing cross-entropy,alleviating class imbalance and improving hard sample performance.
Regarding the core motivational claims in our introduction — including ViT Limitation and four inherent limitations of Mamba-UNet — lack sufficient literature support, we cite references to validate each claim explicitly: a.ViT Limitation:TransUNet[Chen et al.,2021] shows patch-based embedding impairs localization and boundary accuracy. b.Mamba Limitation:LocalMamba[Huang et al.,2024] indicates 2D-to-1D scanning weakens local detail modeling without CNN compensation. c.Skip Connection Limitation:U-Net V2[Peng et al.,2024] verifies naive skip fusion introduces semantic noise, motivating CSAG design. d.Loss Limitation:Focal Dice Loss-based V-Net[Prencipe et al.,2022] confirms standard losses perform poorly on hard samples,justifying our Dice+Focal Loss adoption.
2.Efficiency & Dual-Encoder Cost Justification (R1, R2, R3, Meta) Regarding concerns on model efficiency and computational overhead of the dual-encoder design, we present computational metrics under the same test conditions(Batch Size 1,224×224 input,1×RTX5090,8-class segmentation): Original Mamba-UNet: Params 35.86M | GFLOPs 9.12 | Latency 14.01ms Our VM-NeXT UNet: Params 72.46M | GFLOPs 16.51 | Latency 21.76ms Although parameters increase,GFLOPs remain competitive.21.76ms latency meets clinical real-time requirements(>30 FPS),and this slight tradeoff is worthwhile for a 3.23% DSC improvement from architectural optimization,excluding gains from loss function adjustment.
3.Targeted Ablations (R2, Meta) Regarding the lack of ablation experiments for InjectionBlock and the rationality of loss function selection in ablation studies: a.Injection Block:Isolated ablation results show that without it,the dual-encoder achieves DSC 83.24%,HD95 24.31mm;with it,DSC improves to 83.81% and HD95 decreases to 18.52mm,confirming no redundancy. b.Loss Design:All intermediate ablations adopt Dice+CE Loss,with architecture alone boosting DSC from 82.38% to 85.61%.The final Dice+Focal Loss further improves performance to 86.21%.
4.Response to Inconsistency Between High-Resolution Claim and 224×224 Setup (R2, Meta) Regarding the concern that Mamba’s high-resolution advantage conflicts with our 224×224 input setup: We use 224×224 for fair comparison with TransUNet/SwinUNet.Mamba’s linear complexity supports high-resolution processing without OOM. VMamba [Liu et al.,2024] and VM-UNet [Ruan et al.,2024] confirm SSMs remain competitive at 224×224,with global receptive fields capturing long-range dependencies,validating our claim.
5.Baselines & Visualization (R1, R2, R3) Regarding missing CNN baselines in our evaluation: We have compared our method with CNNs(UNet),recent SOTAs(MERIT, ScaleFormer) and previous MICCAI networks(Reseg-UNet).
Regarding incomplete qualitative visuals and missing CNN visual comparisons: Pure CNNs are well-established baselines,so we focus visuals on advanced models to highlight fusion advantages,while the CNN baseline is included quantitatively.Due to 6-page limit,we only present Synapse visuals;results on Synapse and ACDC verify cross-dataset robustness.
Regarding Synapse dataset configuration: Adopting 8 anatomical categories on Synapse follows standard benchmarks;we will clarify the principle.
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.
I recommend inviting the authors for rebuttal. The reviews are mixed: one reviewer is positive, while two lean weak reject mainly because the contribution appears to be an assembly of existing architectural components rather than a clearly new methodological idea. The rebuttal should focus on clarifying the novelty, adding or explaining efficiency evidence such as parameters, FLOPs, latency, and GPU memory, and justifying the dual-encoder cost. The authors should also address the missing targeted ablations for the injection block, loss design, and other key modules, as well as the mismatch between the high-resolution motivation and the reported 224×224 experiments. A convincing response on these points could materially affect the final decision.
- 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 reading the rebuttal, I recommend Accept, but the paper is close to the threshold. VM-NeXT UNet gives a clear hybrid segmentation architecture: one branch keeps ConvNeXT-style local modeling, the other uses VSS blocks for longer-range context, and the two are connected through injection and gated skip modules. The results on Synapse and ACDC are competitive, and the ablation results support the use of the main components.
The main weakness is novelty. The paper does not introduce a new learning principle; it assembles known pieces from CNN, Mamba/SSM, attention gating, multi-scale decoding, and Dice/Focal training. This makes the contribution more of an architecture design than a methodological advance. The rebuttal does not fully remove this concern, and the final version should avoid claims that are stronger than the evidence.
The efficiency analysis also changes the interpretation of the method. The added numbers show that the model improves accuracy at the cost of more parameters, FLOPs, and latency compared with simpler Mamba-based baselines. This trade-off may still be acceptable, but it should be reported plainly. The paper would also be stronger with cleaner ablations for the loss terms and the injection block, plus multi-run variance.
Overall, I support acceptance because the method is technically coherent, the empirical results are good on both datasets, and the rebuttal adds information needed to judge the architecture. The final version should incorporate the efficiency results, moderate the novelty claims, and clarify which gains come from architecture rather than loss optimization.
Meta-review #2
- 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.
The comments from Reviewer #3 are largely uninformative, which makes the corresponding rating unpersuasive. In contrast, the other reviewers consistently noted that the novelty of the paper is limited. Furthermore, several statements in the rebuttal appear to be misleading:
1.”We use 224×224 for fair comparison with TransUNet/SwinUNet. Mamba’s linear complexity supports high-resolution processing without OOM.”
Since LayerNorm is used, the authors are not constrained by batch-dependent normalization. Therefore, instead of using a batch size of 24 as stated in the paper, they could use a batch size of 1 with a higher input resolution (e.g., 512×512). Given that the GeForce RTX 5090 provides 32GB of memory, it is likely that even a batch size greater than 1 would be feasible.
2.”Although parameters increase, GFLOPs remain competitive. 21.76ms latency meets clinical real-time requirements (>30 FPS) …” (Batch Size 1, 224×224 input, 1×RTX5090, 8-class segmentation)
Since the experiments were conducted on 2D slices extracted from 3D medical images, the claim of real-time clinical applicability is potentially misleading.
Based on these concerns, I recommend rejecting the paper.
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 proposes VM-NeXT UNet, a hybrid U-shaped architecture that combines ConvNeXT and Visual State Space blocks for medical image segmentation. The motivation to improve local detail modeling in Mamba-based designs is sensible, and the reported performance on Synapse and ACDC datasets is competitive.
After rebuttal, two reviewers (R2 and R3) support acceptance, while Reviewer #1 remains negative, citing incremental novelty and the doubled parameter and FLOPs count as weakening the practical value. The efficiency trade-off is a legitimate concern, but Reviewer #2 notes that the accuracy improvements may justify the moderate overhead, and Reviewer #3 considers the ConvNeXT-Mamba synergy a clear methodological contribution. With two of three reviewers favoring acceptance and the performance being well demonstrated, the paper meets the threshold.
I recommend accept. For the camera-ready version, the authors must: (1) incorporate the isolated ablation studies for the injection block and the loss design (BCE, Dice, Focal, and their combinations) to disentangle optimization gains from architectural gains, as requested by Reviewer #2; (2) add statistical significance analysis across multiple runs to confirm the robustness of the reported improvements; (3) explicitly discuss the efficiency trade-off (parameters, FLOPs, latency vs. accuracy) and justify when the added cost is clinically warranted; and (4) integrate all clarifications and analyses from the rebuttal into the main text.