acceptodds
Under review as a conference paper at ICLR 2027

Structure as State, Not Skip: Bridging Semantic Representation and Structural Detail for 3D Medical Image Segmentation

15 views

Abstract

Transformer-based architectures have become widespread in 3D medical image segmentation, yet their gains over strong convolutional baselines remain limited. We trace this mismatch to a representation–structure dilemma: skip connections preserve structure but let information bypass the Transformer; removing them makes the Transformer indispensable but weakens structural detail. The dilemma stems from conflating a task requirement with a U-Net design choice: structure must be preserved, but it need not pass through skip connections. Decoupling the two reveals a third possibility: structure as state, not skip. We realize this principle with Flumina, a multi-stream network that maintains multi-scale persistent states within a shared Vision Transformer backbone. Each sublayer combines these states, applies a shared transformation, and produces token-level updates for each state, while identity residual mappings maintain them across depth. Flumina thus keeps the Transformer responsible for representation learning while preserving structural detail. Experiments across multiple benchmarks and random seeds demonstrate its effectiveness. Code is available at the anonymous link.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
medical image segmentation, vision transformers, medical image analysis, representation learning

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_tSjnOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

Flumina treats features from four CNN tokenizer stages as persistent states within a shared Transformer, rather than sending them to the decoder through separate skip connections. All states are projected onto a common token grid; each sublayer reads a feature-wise combination, applies one shared transformation, and writes updates back through identity residual paths. Final states are restored to their corresponding scales for decoding. The key contribution is a useful organization of structural information throughout representation learning, supported by five segmentation benchmarks and targeted architectural controls.

Critical strengths
  1. An elegant and consequential architectural idea. The paper separates the requirement to preserve structure from the conventional choice of encoder-decoder skips. Distinct scale-grounded states remain available and participate in shared transformations throughout the backbone, while attention operates on one sequence. Although tokenization and multi-stream computation have precedents, their organization here offers a concrete, reusable answer to the structure/representation problem in dense prediction. This conceptual clarity is the main reason for my clear-accept recommendation.

  2. Evidence directly supports the distinctive contribution. Table 5 compares different-scale initialization with replicated states while retaining multiple streams: the proposed initialization improves mean Dice by 1.51 points; it also exceeds early fusion by 2.30 points. Five-task evaluations, three-seed main comparisons, strong baselines including nnU-Net, and measured inference performance reinforce the value beyond an appealing diagram. The control suite makes the central design substantially more convincing.

Critical weaknesses

None that warrants rejection. The residual contraction argument motivates the design but does not prove preservation of anatomical detail in final representations. Small single-seed micro-design differences, such as identity versus mHC, should not be presented as universal superiority. These limitations do not undermine the more substantial contribution of scale-grounded persistent states and its broad empirical support.

Questions for authors

Two clarifications would help calibrate the evidence without requiring a substantially expanded study:

  1. Is the scale-grounded versus replicated-state improvement consistent in any available repeat runs? This is the central comparison; I do not request repeating every micro-design ablation.
  2. Can the main statistical claims distinguish dataset-level effects from pooled-case significance and account for repeated BraTS examinations within patients? This concerns uncertainty estimation, not the patient-disjoint training/test split.

My recommendation is clear acceptance based on the architectural contribution and existing evidence; these questions are intended to refine the scope of the claims.

4 comments
AuthorsOpenReview ↗

Response to Reviewer tSjn (Part 1/2)

We sincerely thank the reviewer for recognizing Flumina’s architectural contribution and for providing constructive feedback. We respond to your questions point by point below.

Q1: Consistency of scale-grounded versus replicated initialization across repeated runs.

Thank you for this suggestion. We acknowledge the importance of repeated runs across random seeds. To make the conclusions of our ablation study more convincing, we have repeated the comparison with the replicated-state variant using multiple random seeds. The results are reported in the table below. After averaging DSC scores over three random seeds, we found that Flumina-B outperformed the replicated-state variant on all five datasets, with an overall mean DSC of 83.64% versus 82.53% across datasets.

VariantAMOS-22LiTSKiTS-23BraTS-25-GLIMAMA-MIAAverage
Flumina-B (3-seed mean)88.2085.6985.8282.0576.4583.64
Replicated-State (3-seed mean)87.87 −0.33\color{#D95360}{-0.33}82.93 −2.76\color{#D95360}{-2.76}85.14 −0.68\color{#D95360}{-0.68}81.44 −0.61\color{#D95360}{-0.61}75.28 −1.17\color{#D95360}{-1.17}82.53 −1.11\color{#D95360}{-1.11}
Replicated-State (Seed 42)87.93 −0.27\color{#D95360}{-0.27}81.00 −4.69\color{#D95360}{-4.69}85.35 −0.47\color{#D95360}{-0.47}81.51 −0.54\color{#D95360}{-0.54}75.61 −0.84\color{#D95360}{-0.84}82.28 −1.36\color{#D95360}{-1.36}
Replicated-State (Seed 1024)87.77 −0.43\color{#D95360}{-0.43}84.88 −0.81\color{#D95360}{-0.81}84.93 −0.89\color{#D95360}{-0.89}81.38 −0.67\color{#D95360}{-0.67}75.46 −0.99\color{#D95360}{-0.99}82.89 −0.75\color{#D95360}{-0.75}
Replicated-State (Seed 2048)87.90 −0.30\color{#D95360}{-0.30}82.89 −2.80\color{#D95360}{-2.80}85.12 −0.70\color{#D95360}{-0.70}81.44 −0.61\color{#D95360}{-0.61}74.76 −1.69\color{#D95360}{-1.69}82.42 −1.22\color{#D95360}{-1.22}

Note. Average is the mean DSC (%) across the five datasets. The highest score in each column is shown in bold. Values in red\color{#D95360}{\textit{red}} indicate decreases from Flumina-B (3-seed mean) in percentage points.

AuthorsOpenReview ↗

Response to Reviewer tSjn (Part 2/2)

Q2: Further details on the statistical analysis.

Per-dataset statistical analysis. Thank you for this suggestion. We have now added per-dataset statistical analyses of DSC, with comparisons between each Flumina variant and the baselines. The table below summarizes the number of statistically significant comparisons. As shown in the table, most comparisons show statistically significant differences, further supporting Flumina’s competitive segmentation performance. We also revised Appendix G.5 to provide detailed statistical results.

VariantAMOS-22LiTSKiTS-23BraTS-25-GLIMAMA-MIATotal
Flumina-S12/12\color{#2E8B57}{12/12}10/12\color{#8A9A32}{10/12}12/12\color{#2E8B57}{12/12}10/12\color{#8A9A32}{10/12}12/12\color{#2E8B57}{12/12}56/60\color{#379130}{56/60}
Flumina-B12/12\color{#2E8B57}{12/12}10/12\color{#8A9A32}{10/12}12/12\color{#2E8B57}{12/12}12/12\color{#2E8B57}{12/12}12/12\color{#2E8B57}{12/12}58/60\color{#2F8E40}{58/60}
Flumina-M12/12\color{#2E8B57}{12/12}10/12\color{#8A9A32}{10/12}12/12\color{#2E8B57}{12/12}12/12\color{#2E8B57}{12/12}12/12\color{#2E8B57}{12/12}58/60\color{#2F8E40}{58/60}
Flumina-L12/12\color{#2E8B57}{12/12}10/12\color{#8A9A32}{10/12}12/12\color{#2E8B57}{12/12}12/12\color{#2E8B57}{12/12}12/12\color{#2E8B57}{12/12}58/60\color{#2F8E40}{58/60}
Total48/48\color{#2E8B57}{48/48}40/48\color{#8A9A32}{40/48}48/48\color{#2E8B57}{48/48}46/48\color{#2F8F3A}{46/48}48/48\color{#2E8B57}{48/48}230/240\color{#2F8F3A}{230/240}

Note. Entries show the number of significant case-level DSC comparisons (p<0.05p<0.05) out of the total tested. Greener shades indicate higher proportions.

Patient-level statistical analysis on BraTS-25-GLI. As detailed in Appendix D, the BraTS-25-GLI test split contains 578 examinations from 366 patients. Following your suggestion, we have averaged the DSC across examinations for each patient and repeated the significance tests at the patient level. The table below report the p-values. The differences between Flumina and most baselines remain statistically significant. For Flumina-M and Flumina-L, all comparisons are statistically significant.

VariantU-NetCoTrSwinUNETRUNETRMedNeXtUNETR++
Flumina-S0.199\quad\color{#D95360}{0.199}0.002\quad\color{#8A9A32}{0.002}0.005\quad\color{#8A9A32}{0.005}<0.001\color{#2E8B57}{<0.001}0.312\quad\color{#D95360}{0.312}<0.001\color{#2E8B57}{<0.001}
Flumina-B0.067\quad\color{#D95360}{0.067}<0.001\color{#2E8B57}{<0.001}0.001\quad\color{#8A9A32}{0.001}<0.001\color{#2E8B57}{<0.001}0.114\quad\color{#D95360}{0.114}<0.001\color{#2E8B57}{<0.001}
Flumina-M0.042\quad\color{#8A9A32}{0.042}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}0.034\quad\color{#8A9A32}{0.034}<0.001\color{#2E8B57}{<0.001}
Flumina-L0.005\quad\color{#8A9A32}{0.005}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}0.017\quad\color{#8A9A32}{0.017}<0.001\color{#2E8B57}{<0.001}
SegMambaZig-RiRSegMamba-V2PrimusCROWnPrimus-V2
Flumina-S0.151\quad\color{#D95360}{0.151}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}
Flumina-B0.030\quad\color{#8A9A32}{0.030}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}
Flumina-M0.018\quad\color{#8A9A32}{0.018}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}
Flumina-L<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}<0.001\color{#2E8B57}{<0.001}

Note. Patient-level pp-values below 0.001 are shown as <0.001<0.001. Green\color{#2E8B57}{\textit{Green}} denotes p<0.001p<0.001, olive\color{#8A9A32}{\textit{olive}} denotes 0.001≤p<0.050.001\leq p<0.05, and red\color{#D95360}{\textit{red}} denotes p≥0.05p\geq0.05.

Thanks for running these extra checks. The one I most wanted to see was whether starting each stream with features from a different scale still helped across multiple runs. Seeing Flumina-B come out ahead on all five datasets in the three-seed averages makes me more confident in that design choice. Averaging the repeat scans for each patient also addresses my concern about the BraTS analysis. I’m keeping my rating at 4.

Dear Reviewer tSjn,

Thank you again for your valuable comments and constructive suggestions. We truly appreciate your support for our work!

Best regards,
Authors of Submission 22518

Reviewer_hPtgOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose solving an intricate issue in 3D Transformer based medical image segmentation architectures, namely the silent skipping of Transformer blocks. This can be witnessed in a number of hybrid architectures containing Transformers and alternative representation learners (eg. convolutions). Their proposed solution is simple - to query internal states derived from the volumetric input via dynamic read and write blocks. They provide comparisons against 12 baselines on the 5 datasets.

Critical strengths
  1. The authors compare against architectures such as MedNeXt, or Primus which are known to perform well, even compared to consistent domain leaders for 3D medical image segmentation such as nnUNet. Their baselines are extensive and strengthen their work.

  2. Their usage of recent advances in the field such as those covered in contemporary work like Primus is encouraging - for example, they layer their changes on recommendations such as not using 16x16x16 tokenizers or the usage of RoPE embeddings.

Critical weaknesses
  1. The obvious weakness is the lack of comparison against the benchmarks in the nnUNet environment. The strength of the framework is the sensible preprocessing, aggressive data augmentation and losses and this benefits the baselines as well. While they do compare against nnUNet itself in G.4, the base architecture is already outperformed by networks such as ResEncL, MedNeXt in literature.

  2. The authors do not mention the design of the decoder in the main paper. As in prior work such as Primus, notes that the usage of a lightweight decoder helps Transformers in learning representations, this might be important to explain.

Questions for authors
  1. Are the authors able to provide any comparisons on just a number of candidate baselines (MedNeXt-L, ResEncL) in the nnUNet framework?

  2. Did the authors experience any challenges with the decoder design, on the representations learned by their Transformer block that might have influenced their choice of decoder?

4 comments
AuthorsOpenReview ↗

Response to Reviewer hPtg (Part 1/2)

We sincerely thank the reviewer for acknowledging the contribution of Flumina and for providing constructive feedback. We respond to these comments point by point below.

W1 & Q1: Limited evaluation within the nnU-Net framework.

Thank you for this suggestion. We recognize the strengths of the nnU-Net framework that you highlighted. To address your concern, we have added results for ResEnc-L and MedNeXt-L within this framework across all five datasets. Since methods such as ResEnc-L typically use larger crop sizes, we have also retrained Flumina-B with the same crop sizes to ensure a fair comparison (denoted as Flumina-B$^{\dagger}$ in the table). The results are summarized in the table below.

MethodAMOS-22-CTAMOS-22-MRILiTSBraTS-25-GLIKiTS-23MAMA-MIAAverage
Flumina-B$^{\dagger}$89.3684.4786.6482.2186.9278.7484.72
Flumina-B88.55 −0.81\color{#D95360}{-0.81}84.56 +0.09\color{#168A65}{+0.09}85.69 −0.95\color{#D95360}{-0.95}82.05 −0.16\color{#D95360}{-0.16}85.82 −1.10\color{#D95360}{-1.10}76.45 −2.29\color{#D95360}{-2.29}83.85 −0.87\color{#D95360}{-0.87}
nnU-Net87.64 −1.72\color{#D95360}{-1.72}83.77 −0.70\color{#D95360}{-0.70}80.23 −6.41\color{#D95360}{-6.41}80.94 −1.27\color{#D95360}{-1.27}82.17 −4.75\color{#D95360}{-4.75}74.20 −4.54\color{#D95360}{-4.54}81.49 −3.23\color{#D95360}{-3.23}
ResEnc-L88.69 −0.67\color{#D95360}{-0.67}84.30 −0.17\color{#D95360}{-0.17}84.73 −1.91\color{#D95360}{-1.91}81.97 −0.24\color{#D95360}{-0.24}85.77 −1.15\color{#D95360}{-1.15}75.85 −2.89\color{#D95360}{-2.89}83.55 −1.17\color{#D95360}{-1.17}
MedNeXt-L88.93 −0.43\color{#D95360}{-0.43}84.29 −0.18\color{#D95360}{-0.18}83.91 −2.73\color{#D95360}{-2.73}82.26 +0.05\color{#168A65}{+0.05}86.57 −0.35\color{#D95360}{-0.35}77.09 −1.65\color{#D95360}{-1.65}83.84 −0.88\color{#D95360}{-0.88}

Note. Average is the mean DSC (%) across the six evaluation columns. The highest score in each column is shown in bold. Differences from Flumina-B$^{\dagger}$ are given in percentage points, with increases in green\color{#168A65}{\textit{green}} and decreases in red\color{#D95360}{\textit{red}}.

AuthorsOpenReview ↗

Response to Reviewer hPtg (Part 2/2)

W2 & Q2: Decoder design of Flumina.

Thank you for this insightful feedback. As detailed in Appendix B, Flumina uses a relatively lightweight decoder. In each stage, trilinear upsampling and the 1×1×11\times1\times1 convolution align spatial resolution and channel dimensions, while a single 3×3×33\times3\times3 convolution spatially refines the fused features.

In our view, some convolutional refinement is needed when decoding multi-scale states. However, consistent with Primus, we also believe that a lightweight decoder may help Transformers learn representations. We therefore use a single 3×3×33\times3\times3 convolution per stage to provide spatial refinement while keeping the decoder lightweight. To assess this design choice, we conducted two additional ablations:

  1. Flumina-B-Lighter: The final 3×3×33\times3\times3 convolution is removed from each decoder stage.
  2. Flumina-B-Heavier: An additional 3×3×33\times3\times3 convolution is added to each decoder stage.

Moreover, a valuable insight from Primus is to avoid spatial interactions in the decoder: its decoder uses only 2×2×22\times2\times2 transposed convolutions. We thus introduce a third variant based on Flumina-B-Lighter:

  1. Flumina-B-Primus: Trilinear upsampling and the 1×1×11\times1\times1 convolution in each decoder stage are replaced with a 2×2×22\times2\times2 transposed convolution.

The results are reported in the table below. We believe these results justify Flumina’s decoder design. In addition, we have revised Section 3.2 to include a full description of the decoder design for completeness.

VariantAMOS-22LiTSKiTS-23BraTS-25-GLIMAMA-MIAAverage
Flumina-B88.1985.3086.3282.1077.0383.79
Flumina-B-Lighter87.78 −0.41\color{#D95360}{-0.41}83.32 −1.98\color{#D95360}{-1.98}85.91 −0.41\color{#D95360}{-0.41}81.58 −0.52\color{#D95360}{-0.52}76.97 −0.06\color{#D95360}{-0.06}83.11 −0.68\color{#D95360}{-0.68}
Flumina-B-Heavier88.21 +0.02\color{#168A65}{+0.02}81.76 −3.54\color{#D95360}{-3.54}85.47 −0.85\color{#D95360}{-0.85}82.23 +0.13\color{#168A65}{+0.13}75.39 −1.64\color{#D95360}{-1.64}82.61 −1.18\color{#D95360}{-1.18}
Flumina-B-Primus88.12 −0.07\color{#D95360}{-0.07}83.47 −1.83\color{#D95360}{-1.83}86.22 −0.10\color{#D95360}{-0.10}82.32 +0.22\color{#168A65}{+0.22}76.21 −0.82\color{#D95360}{-0.82}83.27 −0.52\color{#D95360}{-0.52}

Note. Average is the mean DSC (%) across the five datasets. The highest score in each column is shown in bold. Differences from Flumina-B are given in percentage points, with increases in green\color{#168A65}{\textit{green}} and decreases in red\color{#D95360}{\textit{red}}.

Reviewer_hPtgOpenReview ↗

Thank you for the response

Thank you for your rebuttal. I appreciate the direct response to my questions. Particularly, the architecture results are to my expectations, although I agree that the ResEnc-L's larger patch size introduces a conundrum for fair comparison at times. The authors have addressed it adequately in my opinion. The performance of MedNeXt on normal patch sizes being comparable if not exceeding Flumina-B is also within my expectations.

Nevertheless, this architecture should be available to community for usage and I have accordingly improved my score. I wish the authors all the best.

Dear Reviewer hPtg,

Thank you again for your thoughtful comments and constructive feedback. We sincerely appreciate your support for our work!

Best regards,
Authors of Submission 22518

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

This paper studies the tradeoff between preserving structural details through skip connections and making Transformers learn better representations for 3D medical image segmentation. It proposes Flumina, which uses persistent multi-scale states inside a shared ViT backbone instead of traditional skip connections. Experiments on five medical segmentation benchmarks show improved or competitive Dice scores, with some ablations to study different components.

Critical strengths
  • The motivation is clear. The paper identifies an interesting problem with skip connections, and the idea of keeping structural information as persistent states inside the Transformer is reasonable

  • Experiments cover several datasets and different model sizes. The results generally support the effectiveness of the architecture, and the ablations provide some useful analysis.

Critical weaknesses
  • The novelty is somewhat limited. Multi-scale features, persistent states and residual information propagation are not new ideas. The paper mostly combines them into a new architecture

  • Its not clear whether the improvements really come from removing skip connections, or simply from the additional architectural components. Also the component ablations only use one random seed, which makes the conclusions less convincing.

Questions for authors

1, How does Flumina compare with a strong U-Net/ViT hybrid using better skip feature fusion, under matched parameters and FLOPs?

2, Can the authors repeat the main architectural ablations with multiple random seeds? Are the improvements still consistent?

2 comments
AuthorsOpenReview ↗

Response to Reviewer nycp (Part 1/2)

We sincerely thank the reviewer for acknowledging the contribution of Flumina and for providing constructive feedback. We respond to these comments point by point below.

W1: The novelty is somewhat limited.

Thank you for this comment. Flumina's novelty lies primarily in its architectural design principle, i.e., how it keeps the Transformer responsible for representation learning while preserving structural detail. Our analysis of the representation–structure dilemma shows how skip connections can reduce reliance on Transformer representations. This observation led us to propose a novel architectural design principle: structure as state, not skip. Flumina implements this principle by maintaining multi-scale features as persistent states within a shared ViT. Although some of its modules draw on existing methods, Flumina shows how 3D medical segmentation networks can use structural information throughout representation learning, and we believe this design principle and its implementation can serve as a starting point for futher work.

W2 & Q2: Unclear sources of improvement and single-seed ablations.

Sources of improvement. In fact, our ablation studies help explain where the improvements come from. We briefly summarize the key findings below:

  1. Is the representation–structure dilemma supported by evidence? In semantic encoder–ViT variants, removing skips increases the gains from Transformer blocks and improves class separation, yet lowers overall segmentation performance (Table 4 and Figure 6). These observations provide empirical support for the representation–structure dilemma discussed in the Introduction. Moreover, adding skip connections to Flumina reduces mean Dice by 1.03 points (Table 4).
  2. Do the gains simply come from additional architectural components? The replicated-state variant retains the same architectural components but initializes all streams by replicating the same averaged feature. Flumina improves mean Dice by 1.51 points over this variant (Table 5), supporting the value of our design principle.

We believe these results collectively highlight the potential of our architectural design principle, structure as state, not skip, which also underlies Flumina’s performance improvements.

Single-seed ablations. To make the conclusions of our ablation study convincing, we evaluated the design choices across all five datasets and used mean DSC comparisons. Nevertheless, we also acknowledge the importance of repeated runs across random seeds. To further examine whether the gains come simply from additional architectural components, we have repeated the comparison with the replicated-state variant using multiple random seeds. The results are reported in the table below. After averaging DSC scores over three random seeds, we found that Flumina-B outperformed the replicated-state variant on all five datasets, with an overall mean DSC of 83.64% versus 82.53% across datasets.

VariantAMOS-22LiTSKiTS-23BraTS-25-GLIMAMA-MIAAverage
Flumina-B (3-seed mean)88.2085.6985.8282.0576.4583.64
Replicated-State (3-seed mean)87.87 −0.33\color{#D95360}{-0.33}82.93 −2.76\color{#D95360}{-2.76}85.14 −0.68\color{#D95360}{-0.68}81.44 −0.61\color{#D95360}{-0.61}75.28 −1.17\color{#D95360}{-1.17}82.53 −1.11\color{#D95360}{-1.11}
Replicated-State (Seed 42)87.93 −0.27\color{#D95360}{-0.27}81.00 −4.69\color{#D95360}{-4.69}85.35 −0.47\color{#D95360}{-0.47}81.51 −0.54\color{#D95360}{-0.54}75.61 −0.84\color{#D95360}{-0.84}82.28 −1.36\color{#D95360}{-1.36}
Replicated-State (Seed 1024)87.77 −0.43\color{#D95360}{-0.43}84.88 −0.81\color{#D95360}{-0.81}84.93 −0.89\color{#D95360}{-0.89}81.38 −0.67\color{#D95360}{-0.67}75.46 −0.99\color{#D95360}{-0.99}82.89 −0.75\color{#D95360}{-0.75}
Replicated-State (Seed 2048)87.90 −0.30\color{#D95360}{-0.30}82.89 −2.80\color{#D95360}{-2.80}85.12 −0.70\color{#D95360}{-0.70}81.44 −0.61\color{#D95360}{-0.61}74.76 −1.69\color{#D95360}{-1.69}82.42 −1.22\color{#D95360}{-1.22}

Note. Average is the mean DSC (%) across the five datasets. The highest score in each column is shown in bold. Values in red\color{#D95360}{\textit{red}} indicate decreases from Flumina-B (3-seed mean) in percentage points.

AuthorsOpenReview ↗

Response to Reviewer nycp (Part 2/2)

Q1: Comparison with stronger skip-fusion hybrid methods.

Thank you for this suggestion. In fact, our performance comparisons already include representative methods that improve skip feature fusion. Following the U-shaped paradigm, these methods fall into three main categories:

  1. Enhancing skip features. For instance, SegMamba and SegMamba-V2 introduce FUE to reweight skip features at each scale based on uncertainty.
  2. Improving skip–decoder feature fusion. For example, UNETR++ uses EPA blocks to model spatial and channel dependencies during fusion between skip and decoder features.
  3. Jointly modeling multi-scale features. For example, CoTr stacks deformable Transformer layers after its CNN encoder to jointly model multi-scale features.

Methods in the third category, such as CoTr, may be closest to Flumina in design. However, these methods often add multi-scale modeling as a parallel branch while leaving the original skip connections unchanged. Flumina instead explores organizing structural information as persistent states that are updated throughout representation learning. We believe that results in Tables 2 and 3 support Flumina’s performance advantage over these methods at comparable parameter counts or lower computational cost.

Comments on the paper

1 comment
Area_Chair_xU1ZOpenReview ↗

Discussion Reminder

Dear reviewers, please review the authors’ responses and leave your comments on whether your concerns have been addressed. Thank you!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.