acceptodds
Under review as a conference paper at ICLR 2027

Structured Routing over Propagation Axes and Receptive Supports for Volumetric Medical Image Segmentation

Abstract

Volumetric medical image segmentation requires context that adapts to anatomical direction, spatial location, and semantic class. We introduce VERSA, a structured adaptive routing framework that learns how contextual evidence propagates and which spatial supports inform each prediction. In latent space, V-BAR RWKV learns spatially varying mixtures of bidirectional recurrent responses along three orthogonal axes and integrates them with anisotropic multi-support context. In prediction space, DiVERS constructs nested point, axial, planar-union, and volumetric support experts over decoder features and combines them with boundary-enhanced and pooled-context experts through class- and location-conditioned routing. Normalized expert disagreement regulates the bounded gain of contextual corrections to a point-supported anchor. With 3.67M parameters, VERSA outperforms 19 existing state-of-the-art methods on all 12 public benchmarks spanning CT, MRI, and OCT. Controlled ablations show that learned routing outperforms uniform aggregation with identical candidate axes and support experts, demonstrating the value of adaptive context selection beyond receptive-field expansion.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Volumetric Medical Image Segmentation, Structured Adaptive Routing, Bidirectional Axial Propagation, Receptive-Support Routing

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_7LyGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

VERSA is a compact volumetric segmentation architecture combining spatially routed bidirectional axis-wise RWKV propagation with anisotropic convolutional context. Its prediction head shares convolutional shell responses across nested support experts and combines them through class/spatial routing and disagreement-modulated refinement. The paper evaluates the method across 12 datasets and studies its modules in other architectures. The main contribution is a specific, efficient organization of existing contextual modeling components with broad empirical support.

Critical strengths
  1. Useful and compact architectural integration. The two routing stages address complementary choices: how context propagates and which supports inform a prediction. Shared projections and shell responses keep the model compact at 3.67M parameters, while module insertion experiments support usefulness beyond the proposed backbone.

  2. Broad empirical support. Dice improvements across 12 datasets are reinforced by same-candidate aggregation controls, five-seed comparisons on two datasets, five-fold COVID evaluation, paired case-level statistics, and external LiTS evaluation. These complementary checks support the practical value of the combination, even though not every mechanistic interpretation is fully isolated.

Critical weaknesses

None that presently outweigh the demonstrated architectural and empirical contribution. Uniform-axis controls do not separate spatial adaptation from learned static calibration, and the support-geometry comparison changes parameter sharing. These qualify attribution rather than invalidate the complete system; additional fine-grained ablations are not prerequisites for acceptance.

Questions for authors

Clarifications that could affect the scope of the assessment:

  1. Please state the specific technical distinction from the closest volumetric U-RWKV design discussed in Section 2. The novelty of the concrete routing organization matters more than claiming directional propagation itself is new.
  2. Please clarify validation/checkpoint selection and patient-level partitioning across methods. The existing results support weak acceptance; this question concerns whether the comparative protocol supports that interpretation, not a presumption that it is compromised.
4 comments

1. Please state the specific technical distinction from the closest volumetric U-RWKV design discussed in Section 2. The novelty of the concrete routing organization matters more than claiming directional propagation itself is new.

We thank the reviewer for asking us to clarify this distinction. As shown in Table R16, the key distinction is that U-RWKV relies on fixed, spatially shared tri-axis propagation and fusion, whereas VERSA introduces explicit, input-dependent routing of propagation axes and receptive supports according to spatial location, together with class- and location-conditioned routing of prediction experts.

Table R16: Technical distinction between U-RWKV and VERSA.

AspectU-RWKVVERSA (ours)
Propagation candidatesThree axis-wise WKV scans along D, H, and W.Six bidirectional axis responses (d±d^{\pm}, h±h^{\pm}, w±w^{\pm}), with axis- and direction-specific recurrent dynamics.
Axis aggregationThe three responses are stacked and fused by a shared learned 1×1×11\times1\times1 Conv whose parameters are shared over spatial positions.An input-conditioned spatial simplex router αa(p)\alpha_a(p) produces a location-dependent mixture of the three axis responses.
Local/context routingSpatial Shift and a 3-D convolution; receptive-support selection is implicit.Four explicit anisotropic support bases, including local, dilated, axis-elongated, and coarse context, routed by ηs(p)\eta_s(p).
Cross-path fusionFixed sequential fusion inside the TSE-R block.Recurrent and support paths are fused from their residual updates and their Hadamard-product agreement.
Prediction routingConvolutional segmentation head without explicit support experts.DiVERS routes six support experts with class- and location-conditioned scores; a point anchor is refined by a bounded, disagreement-regulated residual.
Core distinctionTri-axis propagation with spatially shared feature fusion.Hierarchical, spatially and class-conditioned routing over propagation axes and receptive supports.

To verify that this organization matters, we added a direct comparison with U-RWKV (Table R17) on 12 public benchmarks using matched preprocessing, data splits, and evaluation metrics. VERSA improves IoU and Dice on all 12 datasets and reduces HD95 on 11 of 12 datasets. Our controlled ablations (Table 4 and Table 9 in the Manuscript) provide additional evidence for the routing interpretation. These results indicate that the gains arise from adaptive routing and support selection.

Table R17: Quantitative comparison with U-RWKV on 12 public 3D segmentation benchmarks.

MethodVenueMSD Lung Tumour IoUMSD Lung Tumour DiceMSD Lung Tumour HD95MSD Pancreas Tumour IoUMSD Pancreas Tumour DiceMSD Pancreas Tumour HD95MSD Hepatic Vessel IoUMSD Hepatic Vessel DiceMSD Hepatic Vessel HD95
U-RWKVTIP 202652.4467.52107.2353.4767.7910.9048.2462.0335.71
VERSA (Ours)-59.6373.0435.2356.8669.959.6251.2465.8136.59
MethodVenueMSD Colon Cancer IoUMSD Colon Cancer DiceMSD Colon Cancer HD95BreastDCEDL ISPY1 IoUBreastDCEDL ISPY1 DiceBreastDCEDL ISPY1 HD95BreastDCEDL ISPY2 IoUBreastDCEDL ISPY2 DiceBreastDCEDL ISPY2 HD95
U-RWKVTIP 202641.4254.2596.7769.3279.7611.6875.6685.459.99
VERSA (Ours)-43.9657.2853.9775.8184.568.6176.6085.739.64
MethodVenueOIMHS IoUOIMHS DiceOIMHS HD95COVID-19 CT Seg IoUCOVID-19 CT Seg DiceCOVID-19 CT Seg HD95CrossMoDA2021 IoUCrossMoDA2021 DiceCrossMoDA2021 HD95
U-RWKVTIP 202686.9492.652.9570.5781.1758.5267.9179.802.60
VERSA (Ours)-89.7994.402.5873.6183.2315.2270.5081.631.57
MethodVenueParse2022 IoUParse2022 DiceParse2022 HD95ISLES2022 IoUISLES2022 DiceISLES2022 HD95Colorectal-Liver Metastases IoUColorectal-Liver Metastases DiceColorectal-Liver Metastases HD95
U-RWKVTIP 202673.1984.345.2061.8074.316.8777.1185.0711.77
VERSA (Ours)-75.2285.643.8564.5176.804.1778.8186.686.65

2. Please clarify validation/checkpoint selection and patient-level partitioning across methods. The existing results support weak acceptance; this question concerns whether the comparative protocol supports that interpretation, not a presumption that it is compromised.

We thank the reviewer for raising this point. The per-dataset train/test counts are already reported in Table A.1 (Appendix of the manuscript). Here, we clarify how these fixed partitions were used for validation and checkpoint selection. For every method, we used the same fixed patient-level (or subject-level) train/test partition. Within the training partition, we created a deterministic 90/10 train/validation split at the patient/subject level before patch extraction, rather than at the patch or slice level. The resulting patient-level split was identical for VERSA and all baselines in the controlled comparison. We evaluated the validation set every 2,000 Steps and selected the checkpoint with the highest foreground Dice. The selected checkpoint was then frozen and evaluated on the test set. Therefore, all methods (except nnU-Net) in the controlled comparison used identical patient partitions, preprocessing, validation frequency, checkpoint-selection criterion, random seed, and 320,000 Steps training budget. nnU-Net was evaluated using the official nnU-Net protocol while retaining the same patient-level data partition. The complete data split, including the training, validation, and held-out test cases was therefore consistent with that used for VERSA.

We hope that the clarified distinction from U-RWKV and the additional details on patient-level splitting, validation, and checkpoint selection have addressed your remaining questions. If so, we would sincerely appreciate it if you would consider increasing your rating in support of our work.

Thank you for clarifying the distinction from U-RWKV and the patient-level split and checkpoint selection protocol. The direct U-RWKV comparison strengthens the evidence for VERSA as a complete system. The learned static weighting control was especially helpful: it shows that a global axis preference accounts for part of the gain, with spatial adaptation providing a further improvement on the tested task. Together with the four-way ablation, this addresses my main concerns. I remain at weak acceptance, with greater confidence in the contribution of the two-stage design.

Thank you for recognizing our clarifications and the strengthened evidence for VERSA. We are pleased that the direct U-RWKV comparison, learned static weighting control, and four-way ablation addressed your concerns and increased confidence in our design. We sincerely appreciate your thoughtful assessment and constructive feedback.

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

This paper presents VERSA, a framework for volumetric medical image segmentation that adaptively selects contextual information during feature extraction and prediction. V-BAR RWKV combines bidirectional recurrent responses along three orthogonal axes using spatially varying weights and integrates anisotropic multi-support context. DiVERS constructs prediction experts with different geometric supports and combines their outputs through class- and location-conditioned routing. Expert disagreement further regulates the gain of contextual corrections to a point-supported prediction. With 3.67M parameters, VERSA achieves the highest IoU and Dice on all 12 public benchmarks and the lowest HD95 on 10 of them.

Critical strengths

1)This paper addresses context selection at two distinct stages of volumetric segmentation. V-BAR RWKV adaptively combines axis-wise recurrent responses, while DiVERS selects among geometrically defined prediction supports based on class and spatial location. The four nested supports in Section 3.3 provide a clear geometric formulation that connects the prediction mechanism to the paper's motivation. 2)The paper provides extensive experimental evaluation across 12 benchmarks, with VERSA achieving the highest IoU and Dice on all of them in Table 1. Tables 3 and 7 show consistent improvements when V-BAR RWKV and DiVERS are separately incorporated into eight segmentation architectures. Tables 4-6 and 8-10 further examine the contributions of individual components and compare them with alternative designs. In particular, Table 9 provides targeted evaluations of point-anchored refinement, class-conditioned routing, and disagreement regulation.

Critical weaknesses

1)Section 4.4 ablates V-BAR RWKV and DiVERS in separate experiments but does not compare the backbone, each module individually, and their combination under the same experimental setting. This makes it difficult to assess the complementary contributions of the two routing stages. In addition, Tables 4 and 5 compare learned routing with uniform aggregation but do not include static learnable fusion baselines, leaving the specific benefit of spatially adaptive weighting unclear. 2)Section 2 discusses U-RWKV, which also employs tri-axis WKV scanning and learned feature fusion, but this closely related method is absent from Tables 1 and 6. Although the paper describes additional spatial routing and fusion mechanisms in V-BAR RWKV, the lack of a direct comparison makes it difficult to assess their practical advantages over this existing approach.

Questions for authors

1)Section 4.4 examines V-BAR RWKV and DiVERS separately. Can the authors provide a comparison of the backbone, V-BAR alone, DiVERS alone, and full VERSA on the same dataset under identical training conditions to clarify the individual and combined contributions of the two modules? 2)Tables 4 and 5 compare learned routing with uniform aggregation. How does spatially adaptive routing compare with static learnable weighting over the same candidate axes and supports? 3)Section 2 discusses U-RWKV, which also employs tri-axis WKV scanning and learned feature fusion, but it is absent from Tables 1 and 6. Can the authors explain this omission and provide a direct comparison under the same evaluation protocol? 4)Can the authors clarify the implementation of hybrid pooling and the decoder gates in Figure 2, as well as the specific architectures and output dimensions of the routing operators in Eqs. (4)-(6)? How is the coarse contextual response (B3B_3) computed? 5)What motivates the use of mean and maximum absolute expert logits as global routing evidence in Eq. (9)? What is the rationale for using normalized expert disagreement to regulate correction strength in Eqs. (10)-(11)? 6)In Table 8, the random-partition and geometric expert variants differ in both support geometry and parameter sharing. Can the authors isolate the contribution of support geometry through a comparison with matched parameter sharing?

7 comments

1. Section 4.4 examines V-BAR RWKV and DiVERS separately. Can the authors provide a comparison of the backbone, V-BAR alone, DiVERS alone, and full VERSA on the same dataset under identical training conditions to clarify the individual and combined contributions of the two modules?

We thank the reviewer for this suggestion. We provide a controlled four-way ablation on MSD Lung Tumour, MSD Hepatic Vessel, and OIMHS (Table R8). All variants use the same training and evaluation protocol, while all components other than V-BAR RWKV and DiVERS are kept fixed. Specifically, the Backbone contains neither module, the two intermediate variants contain exactly one module, and Full VERSA contains both modules. Both modules are individually effective. The joint use of both modules gives the best IoU, Dice, and HD95 on all three datasets. Thus, the consistent additional gains over either single-module variant suggest that V-BAR RWKV and DiVERS provide complementary contributions rather than redundant improvements. We have added this controlled comparison.

Table R8: Controlled four-way ablation under identical training and evaluation settings.

VariantMSD Lung Tumour — IoUMSD Lung Tumour — DiceMSD Lung Tumour — HD95MSD Hepatic Vessel — IoUMSD Hepatic Vessel — DiceMSD Hepatic Vessel — HD95OIMHS — IoUOIMHS — DiceOIMHS — HD95
Backbone54.8968.5261.0648.8363.2843.3787.6192.962.80
Backbone + V-BAR RWKV56.7871.4556.5750.2064.1438.1888.2893.482.71
Backbone + DiVERS57.0571.0352.7550.3964.3137.5788.4593.502.70
Full VERSA59.6373.0435.2351.2465.8136.5989.7994.402.58

2. Tables 4 and 5 compare learned routing with uniform aggregation. How does spatially adaptive routing compare with static learnable weighting over the same candidate axes and supports?

We thank the reviewer for raising this point. We conducted controlled ablations in which the candidate propagation axes/supports are kept fixed within each comparison, and only the routing parameterization changes (Table R9 and R10). Specifically, uniform aggregation uses fixed equal coefficients; the static baseline learns one global weighting vector, Softmax⁡(gd,gh,gw)\operatorname{Softmax}(g_d,g_h,g_w) or Softmax⁡(h0,…,h3)\operatorname{Softmax}(h_0,\ldots,h_3), which is shared across all spatial locations; our method instead predicts spatially varying normalized weights, αa(p)\alpha_a(p) or ηs(X,p)\eta_s(X,p). The static learnable baseline already improves over uniform aggregation, confirming that learning the relative importance of candidate axes/supports is beneficial. However, it can only represent a dataset-level average preference. Spatially adaptive routing further improves all metrics over static weighting.

Table R9: Controlled ablation of propagation-axis routing in V-BAR RWKV on MSD Pancreas Tumour.

VariantRouting weightsLearnable?Spatially varying?IoUDiceHD95
Uniform axis aggregation(13,13,13)\left(\frac{1}{3},\frac{1}{3},\frac{1}{3}\right)×\times×\times54.5268.1311.18
Global learned static axis weightingSoftmax⁡(gd,gh,gw)\operatorname{Softmax}(g_d,g_h,g_w)✓\checkmark×\times56.1569.479.86
Spatially adaptive axis routing (Ours)αa(p)\alpha_a(p)✓\checkmark✓\checkmark56.8669.959.62

Table R10: Controlled ablation of receptive-support routing in V-BAR RWKV on Colorectal-Liver Metastases.

VariantRouting weightsLearnable?Spatially varying?IoUDiceHD95
Equal-weight support averaging(14,14,14,14)\left(\frac{1}{4},\frac{1}{4},\frac{1}{4},\frac{1}{4}\right)×\times×\times77.5485.6810.24
Global learned static support weightingSoftmax⁡(h0,h1,h2,h3)\operatorname{Softmax}(h_0,h_1,h_2,h_3)✓\checkmark×\times78.0386.059.84
Spatially adaptive support routing (Ours)ηs(X,p)\eta_s(X,p)✓\checkmark✓\checkmark78.8186.686.65

Since the static and spatially adaptive variants use the same candidate axes/supports, these additional gains isolate the benefit of spatial conditioning rather than a larger candidate set. A static learnable vector captures only the global average importance of each propagation direction or receptive support, whereas spatially adaptive routing can select different directions or support scales according to local feature context, object extent, and boundary structure. The particularly large HD95 improvement for receptive-support routing further suggests better handling of difficult boundaries and local outlier errors.

3. Section 2 discusses U-RWKV, which also employs tri-axis WKV scanning and learned feature fusion, but it is absent from Tables 1 and 6. Can the authors explain this omission and provide a direct comparison under the same evaluation protocol?

We thank the reviewer for pointing this out. We have now added U-RWKV as a complete baseline and re-evaluated it under the same data splits, preprocessing, training/evaluation protocol, and metrics as VERSA. Across all 12 public 3D segmentation benchmarks, VERSA achieves higher IoU and Dice than U-RWKV on all 12 datasets, with lower HD95 on 11 out of 12 datasets (Table R11). Table 6 in the manuscript is designed for module-level replacement within the fixed VERSA framework, whereas U-RWKV is a complete segmentation architecture. To further provide a controlled module-level comparison, we have added an ablation by replacing V-BAR RWKV with TSE-R (Table R12), a key component of U-RWKV, while keeping all other components and the evaluation protocol unchanged. The TSE-R variant obtains 63.69 IoU, 76.19 Dice, and 5.40 HD95, compared with 64.51, 76.8076.80, and 4.17 for V-BAR RWKV. Thus, V-BAR RWKV improves IoU and Dice by 0.82 and 0.61 points, respectively, and reduces HD95 by 1.23. These results support the effectiveness of the proposed V-BAR RWKV design beyond the shared use of tri-axis WKV scanning and learned feature fusion.

Table R11: Quantitative comparison with U-RWKV on 12 public 3D segmentation benchmarks.

MethodVenueMSD Lung Tumour IoUMSD Lung Tumour DiceMSD Lung Tumour HD95MSD Pancreas Tumour IoUMSD Pancreas Tumour DiceMSD Pancreas Tumour HD95MSD Hepatic Vessel IoUMSD Hepatic Vessel DiceMSD Hepatic Vessel HD95
U-RWKVTIP 202652.4467.52107.2353.4767.7910.9048.2462.0335.71
VERSA (Ours)–59.6373.0435.2356.8669.959.6251.2465.8136.59
MethodVenueMSD Colon Cancer IoUMSD Colon Cancer DiceMSD Colon Cancer HD95BreastDCEDL ISPY1 IoUBreastDCEDL ISPY1 DiceBreastDCEDL ISPY1 HD95BreastDCEDL ISPY2 IoUBreastDCEDL ISPY2 DiceBreastDCEDL ISPY2 HD95
U-RWKVTIP 202641.4254.2596.7769.3279.7611.6875.6685.459.99
VERSA (Ours)–43.9657.2853.9775.8184.568.6176.6085.739.64
MethodVenueOIMHS IoUOIMHS DiceOIMHS HD95COVID-19 CT Seg IoUCOVID-19 CT Seg DiceCOVID-19 CT Seg HD95CrossMoDA2021 IoUCrossMoDA2021 DiceCrossMoDA2021 HD95
U-RWKVTIP 202686.9492.652.9570.5781.1758.5267.9179.802.60
VERSA (Ours)–89.7994.402.5873.6183.2315.2270.5081.631.57
MethodVenueParse2022 IoUParse2022 DiceParse2022 HD95ISLES2022 IoUISLES2022 DiceISLES2022 HD95Colorectal-Liver Metastases IoUColorectal-Liver Metastases DiceColorectal-Liver Metastases HD95
U-RWKVTIP 202673.1984.345.2061.8074.316.8777.1185.0711.77
VERSA (Ours)–75.2285.643.8564.5176.804.1778.8186.686.65

Table R12: Comparison of V-BAR RWKV with alternative modules in VERSA.

Replacement moduleIoUDiceHD95Replacement moduleIoUDiceHD95
SE64.0976.294.28CBAM63.2175.166.00
ResConv3D62.4675.006.61ConvNeXt3D64.2276.114.70
Transformer63.8775.824.98Swin Transformer64.3476.334.27
Efficient Paired Attention64.3376.314.44Mamba layer64.1776.174.19
Vision Mamba63.8675.874.47RWKV63.1274.965.66
Vision-RWKV63.8176.004.73TSE-R63.6976.195.40
V-BAR RWKV (Ours)64.5176.804.17

4. Can the authors clarify the implementation of hybrid pooling and the decoder gates in Figure 2, as well as the specific architectures and output dimensions of the routing operators in Eqs. (4)-(6)? How is the coarse contextual response (B3B_3) computed?

We thank the reviewer for pointing this out. Unless otherwise stated, all tensors use the channel-first format [B,C,D,H,W][B,C,D,H,W].

Hybrid pooling and decoder gates. At each of the three encoder downsampling stages, we use a channel-wise learnable convex combination of 33D max pooling and average pooling, both with kernel size 232^3 and stride 22. The mixing coefficient is obtained from a learnable per-channel sigmoid parameter, initialized such that the initial contribution of max and average pooling is 0.50.5 each. The mixed feature is then processed by a residual 3×3×33\times3\times3 convolutional block, which also changes the channel width.

In each decoder block, the upsampled decoder feature and the aligned encoder skip feature are first mapped to the same number of channels, denoted by CoC_o. A spatial gate is predicted from the concatenation of the decoder feature, the skip feature, and their absolute difference. It consists of a 1×1×11\times1\times1 convolution, GroupNorm, SiLU, and a 3×3×33\times3\times3 convolution, and produces a spatial map of size [B,1,D,H,W][B,1,D,H,W]. A channel gate is produced from the global average pooled decoder--skip concatenation using two 1×1×11\times1\times1 convolutions and SiLU, yielding a channel vector of size [B,Co,1,1,1][B,C_o,1,1,1]. The gated skip feature is added to the decoder feature. We also inject a separately scaled high-frequency skip residual obtained by subtracting a stride-one 3×3×33\times3\times3 average-pooled skip feature, followed by a residual refinement block. For the default configuration, the decoder output widths are 9696, 4848, and 2424.

Routing operators in Eqs.~(4)--(6). The three routing operators are lightweight pointwise convolutional routers; their outputs are normalized over the candidate branches by softmax.

**Eq.~(4), axis router. The axis router is a 1×1×11\times1\times1 convolution--SiLU--1×1×11\times1\times1 convolution module mapping CrC_r channels to hah_a hidden channels and then to three outputs, corresponding to the depth, height, and width axes. Here ha=max⁡(16,⌊Cr/4⌋)h_a=\max(16,\lfloor C_r/4\rfloor). The router is evaluated on a grid with at most 636^3 cells, and the resulting three-way weight map is trilinearly interpolated to the recurrent grid. With the default Cr=96C_r=96, ha=24h_a=24, and a 12312^3 bottleneck grid, the output has shape [B,3,12,12,12][B,3,12,12,12]. Each routed axis response has shape [B,96,12,12,12][B,96,12,12,12]. The local--axis gate is a learned per-channel sigmoid vector broadcast over the spatial dimensions.

Eq.~(5), support router. The anisotropic context bridge first reduces the bottleneck feature to CiC_i channels and constructs four support responses: local, dilated, axis-elongated, and coarse-context responses. The support router is a single 1×1×11\times1\times1 convolution mapping CiC_i channels to four outputs. In the default configuration, Ci=48C_i=48, and the routing weights have shape [B,4,12,12,12][B,4,12,12,12]. The four weighted responses remain 4848-channel features before being projected back to the bottleneck width C=192C=192.

Eq.~(6), path-fusion router. The path router receives the concatenation of the context update, the RWKV update, and their element-wise product. It uses a 1×1×11\times1\times1 convolution, GroupNorm, SiLU, and a final 1×1×11\times1\times1 convolution to map 3C3C channels to two outputs. With C=192C=192, the architecture is 576→24→2576\rightarrow24\rightarrow2, producing a two-way routing map of shape [B,2,12,12,12][B,2,12,12,12] for the default bottleneck. The cross-interaction branch uses a grouped 3×3×33\times3\times3 convolution followed by GroupNorm, SiLU, and a 1×1×11\times1\times1 convolution, while preserving the channel width.

Coarse contextual response B3B_3. The coarse response is not obtained by global average pooling to a single vector. The bottleneck feature is first projected and normalized to CiC_i channels. It is then adaptively average-pooled to at most 434^3 cells, processed by a grouped 3×3×33\times3\times3 convolution, GroupNorm, SiLU, and a 1×1×11\times1\times1 convolution, and finally trilinearly upsampled back to the bottleneck resolution. Thus, B3B_3 is a learned coarse spatial feature map with the same spatial dimensions as the other support responses, rather than a global scalar or vector.

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

This paper proposed VERSA for 3D medical image segmentation. The main idea is to use the adaptive routing to select most contextual information from different propagation directions and receptive supports. The proposed method contains RWKV for spatial axis routing and DiVERS for class-specific receptive support selection. Experiments on multiple public datasets demonstrated the effectiveness of the proposed method.

Critical strengths
  1. Context modelling is important for 3D medical image segmentation, especially for anatomical structures with different shapes and orientations. This paper proposed an interesting framework to adaptively select the spatial context rather than using fixed propagation directions.
  2. The experiment results are also promising. The proposed method achieved competitive results on 12 public benchmarks compared with many baseline methods. The performance across CT, MRI, and OCT datasets is encouraging.
  3. The ablation studies are relatively comprehensive. Especially, the comparisons between learned routing and uniform aggregation are meaningful to demonstrate the effectiveness of the proposed routing strategy.
Critical weaknesses
  1. My main concern is the very complicated framework. The method introduces multiple routing strategies, recurrent operations, receptive support experts, and disagreement-based modulation. Although the results are promising, it is unclear whether all these components are really necessary.
  2. Although the motivation is interesting, the current analysis is not sufficient to explain how the learned routing is related to anatomical structures. E.g., the authors claimed that different anatomical regions require different propagation directions. However, it is not clear whether the learned routing really follows these anatomical characteristics.
  3. Many baselines are included in the experiment. However, i did not see the most important baseline nnUNet. To my understanding, the main concerns raised by the authors have been effectively solved by nnUNet frameworking with self-adapted hyper-parameters.
  4. Although the model has relatively few parameters, this does not necessarily mean that it is computationally efficient. The inference speed and computational cost should be discussed more carefully.
Questions for authors
  1. Can you clarify the anatomical meaning of the learned routing and explain how the current experiments support the claimed adaptive context selection?
  2. Please clarify the differences and complementary roles of V-BAR RWKV and DiVERS, and discuss the limitations of fixed-axis routing.
  3. Add the nnUNet as baseline. Otherwise, the improvement is not convincing at all.
7 comments

1. Can you clarify the anatomical meaning of the learned routing and explain how the current experiments support the claimed adaptive context selection?

We thank the reviewer for raising this point. We clarify that the anatomical meaning of routing is geometric and evidential. V-BAR RWKV computes spatially varying weights over three fixed D/H/WD/H/W propagation axes. A larger weight indicates that information transported along the corresponding image axis is more useful at that location. This preference may reflect both target continuity and sampling geometry. DiVERS performs soft, class- and location-conditioned reweighting over six fixed supports: point, axial, planar-union, volumetric, boundary-enhanced, and pooled-context. The point-supported logit acts as an anchor, while broader supports provide bounded contextual corrections. The added statistics are internally consistent: the weights within each V-BAR RWKV and DiVERS block sum to approximately 100% (Table R1). V-BAR RWKV favors the WW axis for the Parse2022 pulmonary artery (52.04%) and the DD axis for the BreastDCEDL ISPY2 tumor (96.09%). DiVERS favors pooled context for the liver (47.95%), axial context for liver metastases (41.02%), and axial plus planar supports for the breast tumor (39.75% and 32.61%). Because V-BAR RWKV is class-agnostic, the liver and liver-metastasis rows are ground-truth-stratified summaries of the same spatial router. The positive boundary--interior TV intervals in Table R2 indicate location-dependent routing.

Table R1: Routing weights for V-BAR RWKV and DiVERS.

DatasetTarget labelV-BAR RWKV: DV-BAR RWKV: HV-BAR RWKV: WDiVERS: PointDiVERS: AxialDiVERS: Planar-unionDiVERS: VolumetricDiVERS: Boundary-enhancedDiVERS: Pooled-context
Colorectal–Liver MetastasesLiver56.85%24.27%18.88%3.65%12.93%10.88%21.11%3.49%47.95%
Colorectal–Liver MetastasesLiver metastases51.17%28.76%20.07%11.54%41.02%5.77%8.58%10.97%22.12%
Parse2022Pulmonary artery25.01%22.95%52.04%2.28%16.98%24.47%8.25%2.43%45.58%
BreastDCEDL ISPY2Tumor96.09%0.87%3.04%0.92%39.75%32.61%13.64%2.52%10.57%

Table R2: Boundary–interior routing difference measured by total variation. Values are reported as mean TV with a 95% bootstrap confidence interval.

DatasetTarget labelV-BAR RWKV three-axis TVDiVERS six-expert TV
Colorectal–Liver MetastasesLiver0.0931; [0.0881, 0.0980]0.0498; [0.0410, 0.0584]
Colorectal–Liver MetastasesLiver metastases0.0235; [0.0192, 0.0280]0.1104; [0.0997, 0.1214]
Parse2022Pulmonary artery0.2081; [0.1911, 0.2230]0.1212; [0.1097, 0.1339]
BreastDCEDL ISPY2Tumor0.0294; [0.0274, 0.0317]0.0884; [0.0804, 0.0968]

The stronger evidence comes from matched ablations with identical candidate contexts. On MSD Pancreas Tumour, learned V-BAR RWKV routing improves over uniform axis aggregation from 54.52/68.13/11.18 to 56.86/69.95/9.62 in IoU/Dice/HD95 (Table 4 in the manuscript). On Colorectal–Liver Metastases, learned support weighting improves over equal-weight averaging from 77.54/85.68/10.24 to 78.81/86.68/6.65 (Table 5 in the manuscript). For DiVERS, full routing improves over uniform routing on CrossMoDA from 69.84/81.04/1.68 to 70.50/81.63/1.57 (Table 9 in the manuscript). These results support adaptive context selection beyond simply enlarging the receptive field.

2. Please clarify the differences and complementary roles of V-BAR RWKV and DiVERS, and discuss the limitations of fixed-axis routing.

We thank the reviewer for raising this point. V-BAR RWKV and DiVERS address different conditional decisions at different stages of the network (Table R3).

V-BAR RWKV answers “along which propagation axis should information travel?”, whereas DiVERS answers “which receptive support should contribute to this class prediction, and by how much?”. The two modules are therefore complementary: V-BAR RWKV transports and enriches contextual features before decoding, while DiVERS performs class- and location-specific support selection at the prediction stage.

Table R3: Comparison between V-BAR RWKV and DiVERS.

AspectV-BAR RWKVDiVERS
Operating stageEncoder bottleneckDecoder prediction space
Conditional decisionAlong which propagation axis should context travel?Which receptive support should contribute to each class prediction?
Routed objectBidirectional recurrent responses along depth, height, and width axes, together with anisotropic context responsesSix support experts: point, axial, planar-union, volumetric, boundary-enhanced, and pooled-context experts
Routing signalSpatially varying axis weights, with direction and update-conditioned fusionClass-conditioned global scores, class-shared spatial scores, and disagreement-regulated residual gain
Aggregation formFeature-level contextual propagation and fusionPoint-anchored, bounded residual refinement of class logits
Main purposeTransport long-range, direction-aware latent features and model anatomical continuitySelect the appropriate support scale and shape for each class and voxel while preserving boundaries

To isolate their complementary utility, we added a module-level ablation in Table R4 using the same data split, optimization schedule, crop size, inference protocol, and random seed. Removing either module degrades the overlap metrics, while the full model achieves the highest IoU and Dice and the lowest HD95 on all three datasets. Overall, these results support complementary contributions.

Table R4: Controlled four-way ablation under identical training and evaluation settings.

VariantMSD Lung Tumour — IoUMSD Lung Tumour — DiceMSD Lung Tumour — HD95MSD Hepatic Vessel — IoUMSD Hepatic Vessel — DiceMSD Hepatic Vessel — HD95OIMHS — IoUOIMHS — DiceOIMHS — HD95
Backbone54.8968.5261.0648.8363.2843.3787.6192.962.80
Backbone + V-BAR RWKV56.7871.4556.5750.2064.1438.1888.2893.482.71
Backbone + DiVERS57.0571.0352.7550.3964.3137.5788.4593.502.70
Full VERSA59.6373.0435.2351.2465.8136.5989.7994.402.58

We thank the reviewer for requesting clarification on fixed-axis routing. To disentangle weight learnability from spatial adaptivity, we added a new global learned static axis weighting baseline. As shown in Table R5, global learned weighting improves over uniform aggregation. This indicates that learning a dataset-level axis preference is already more effective than assigning equal weight to all axes. The full spatially adaptive router further improves the results. In addition, all three single-axis bidirectional variants underperform the full model, supporting the complementary utility of D-, H-, and W-wise context.

These results isolate two distinct effects: learning the relative importance of the candidate axes improves over uniform aggregation, while allowing this preference to vary with the input and spatial location provides an additional benefit beyond a single global axis prior.

Table R5: Controlled ablation of propagation-axis routing in V-BAR RWKV on MSD Pancreas Tumour.

VariantRouting weightsLearnable?Spatially varying?IoUDiceHD95
Depth-only bidirectional propagation(1,0,0)××54.5867.7010.31
Height-only bidirectional propagation(0,1,0)××55.6968.6410.25
Width-only bidirectional propagation(0,0,1)××55.1168.1010.09
Uniform axis aggregation(13,13,13)\left(\frac{1}{3},\frac{1}{3},\frac{1}{3}\right)××54.5268.1311.18
Global learned static axis weightingSoftmax⁡(gd,gh,gw)\operatorname{Softmax}(g_d,g_h,g_w)✓×56.1569.479.86
Spatially adaptive axis routing (Ours)αa(p)\alpha_a(p)✓✓56.8669.959.62

We also acknowledge the remaining limitation. Although the routing weights are spatially adaptive, the candidate propagation basis is still restricted to the predefined D/H/W axes. Locally rotated, spacing-aware, or deformable propagation bases are promising extensions.

3. Add the nnUNet as baseline. Otherwise, the improvement is not convincing at all.

Thank you for pointing this out. We agree that nnU-Net (Isensee et al., Nature Methods, 2021) is an essential and strong reference baseline for evaluating 3D medical image segmentation methods. nnU-Net was evaluated using the official nnU-Net protocol while retaining the same patient-level data partition.

In the direct comparison, VERSA achieves the highest IoU and Dice on every benchmark (Table R6). It also obtains the lowest HD95 on all but one benchmark. These results demonstrate that VERSA consistently outperforms nnU-Net in regional overlap and achieves better boundary accuracy in nearly all cases.

These results show that VERSA’s gains remain consistent against a strong canonical baseline across heterogeneous CT, MRI, and OCT benchmarks.

Table R6: Quantitative comparison with nnU-Net on 12 public 3D segmentation benchmarks.

MethodVenueMSD Lung Tumour — IoUMSD Lung Tumour — DiceMSD Lung Tumour — HD95MSD Pancreas Tumour — IoUMSD Pancreas Tumour — DiceMSD Pancreas Tumour — HD95MSD Hepatic Vessel — IoUMSD Hepatic Vessel — DiceMSD Hepatic Vessel — HD95
nnU-NetNature methods 202157.5570.5061.2256.5069.7611.6950.9865.4834.22
VERSA (Ours)-59.6373.0435.2356.8669.959.6251.2465.8136.59
MethodVenueMSD Colon Cancer — IoUMSD Colon Cancer — DiceMSD Colon Cancer — HD95BreastDCEDL ISPY1 — IoUBreastDCEDL ISPY1 — DiceBreastDCEDL ISPY1 — HD95BreastDCEDL ISPY2 — IoUBreastDCEDL ISPY2 — DiceBreastDCEDL ISPY2 — HD95
nnU-NetNature methods 202143.5456.93166.9072.6182.8319.2274.8384.5511.55
VERSA (Ours)-43.9657.2853.9775.8184.568.6176.6085.739.64
MethodVenueOIMHS — IoUOIMHS — DiceOIMHS — HD95COVID-19 CT Seg — IoUCOVID-19 CT Seg — DiceCOVID-19 CT Seg — HD95CrossMoDA2021 — IoUCrossMoDA2021 — DiceCrossMoDA2021 — HD95
nnU-NetNature methods 202188.7393.842.8870.8181.1831.8270.0081.051.62
VERSA (Ours)-89.7994.402.5873.6183.2315.2270.5081.631.57
MethodVenueParse2022 — IoUParse2022 — DiceParse2022 — HD95ISLES2022 — IoUISLES2022 — DiceISLES2022 — HD95Colorectal-Liver Metastases — IoUColorectal-Liver Metastases — DiceColorectal-Liver Metastases — HD95
nnU-NetNature methods 202173.8584.804.3763.9476.135.8678.1286.2512.75
VERSA (Ours)-75.2285.643.8564.5176.804.1778.8186.686.65

4. Although the model has relatively few parameters, this does not necessarily mean that it is computationally efficient. The inference speed and computational cost should be discussed more carefully.

We thank the reviewer for pointing this out. We report parameter count, FLOPs, and measured inference throughput in Table R7. Using a (96×96×96)(96\times96\times96) input patch, batch size 1, and one NVIDIA RTX 4090, VERSA requires 3.67M parameters and 107.67 GFLOPs per patch, while achieving 37.39 patches/s. These results support a favorable compactness–compute–speed trade-off.

Table R7: Comparison of model parameters, FLOPs, and inference throughput.

MethodParams (M)FLOPs (G)Throughput (patches/s)
3D U-Net5.75135.8875.25
nnU-Net27.66187.8763.38
UNETR92.6282.5858.98
SwinUNETR61.99329.4620.37
3D UX-Net53.00631.7618.83
nnFormer149.10224.3652.60
MedNeXt11.64174.109.78
UNesT84.88258.1139.86
SegMamba64.24655.6219.03
SwinSMT15.5183.7128.95
VSmTrans49.86358.8915.93
SuperLightNet2.7519.5238.36
WaveFormer16.86371.3818.64
SlimFormer3D17.0278.3555.71
MambaHoME170.1654.1012.29
SlimUNETR V222.0354.75116.92
LightUNETR-L3.9354.3329.79
Rep3D65.83766.603.82
MMSegRWKV37.28282.8824.27
HCRT6.2674.6457.91
U-RWKV9.2392.3728.41
VERSA (Ours)3.67107.6737.39

We hope that the additional controlled ablations, nnU-Net comparison, efficiency analysis, and clarification of the geometric and anatomical interpretation have addressed your main concerns. If so, we would sincerely appreciate it if you would consider increasing your rating in support of our work.

Comments on the paper

1 comment
Area_Chair_mEx5OpenReview ↗

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.