acceptodds
Withdrawn from ICLR 2027

SPARNet: Breaking the Detail–Direction Trade-off in 3D Medical State-Space Models with Scale-Preserving Radial Scanning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

State-space models (SSMs) have emerged as an efficient alternative to Transformers for 3D medical image segmentation, yet how volumetric features are tokenized and serialized for these models remains largely overlooked. We identify two coupled limitations in existing 3D SSMs. First, fixed and coarse token scales cause detail loss at the boundaries of small or low-contrast structures. Second, Cartesian-axis serialization separates neighboring voxels and introduces directional bias into state propagation. Because tokenization determines what information is preserved and scan order determines how it propagates, we argue that the two must be addressed jointly. We present SPAR-Net, a Scale-Preserving Adaptive Radial Network built on this principle. A Scale-Adaptive Feature Extractor aligns fine-, intermediate, and coarse-scale features on a shared high-resolution token grid, which adapts the receptive field without discarding spatial detail. Stochastic Layer Scanning replaces axis-aligned scans with center-referenced radial layers traversed along complementary inward and outward paths. A Feature Refinement block restores volumetric correspondence after sequence modeling, and a Bidirectional Progressive Fusion Block exchanges semantic and boundary cues across encoder levels. Under a unified protocol, SPARNet consistently improves both Dice and HD95 over existing 3D segmentation models. It achieves 85.25% mean Dice and 12.20 HD95 on Synapse multi-organ CT, and 64.02% mean Dice and 17.15 HD95 on MSD Pancreas. The largest gains appear on boundary-sensitive organs such as the gallbladder and pancreas, and on pancreatic tumors. Analyses of token scale, scan order, feature fusion, and efficiency show that every component contributes. They also show that fine multi-scale tokenization and radial bidirectional propagation are complementary.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
State-Space Models, 3D Medical Image Segmentation, Radial Scanning, Multi-Scale Tokenization, Directional Bias

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_EaPUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SPARNet is a 3D medical segmentation architecture intended to address detail loss from coarse tokenization and directional bias from Cartesian serialization in state-space models. SAFE aligns multiple patch scales on a fine token grid; SLS uses crop-centered radial layers with complementary inward/outward traversals; feature refinement and bidirectional fusion restore and combine volumetric features. Evaluation uses Synapse abdominal CT and MSD Pancreas CT under a shared training protocol. The eight-region rotational variant reports 85.25% mean Dice / 12.20 HD95 on Synapse and 64.02% / 17.15 on MSD Pancreas. Component, scan-order, layer-width, and plug-in ablations investigate the design.

Critical strengths
  1. A useful investigation of serialization for volumetric state-space models. The radial scanning design provides a concrete alternative to axis-aligned flattening, and the scan-order ablation holds the surrounding backbone fixed. The supplementary plug-in experiments with U-Mamba, SegMamba, and SegMamba-V2 report HD95 improvements without changing parameter counts, providing evidence that the scan itself has value beyond the full architecture. These experiments are more informative than a comparison of complete networks alone.

  2. Promising boundary-sensitive results with several relevant controls. The eight-region variant improves both reported mean Dice and HD95 over the evaluated baselines on the two datasets, with particularly encouraging pancreatic tumor results. The paper includes organ-level results, repeated Synapse runs, scale/fusion ablations, and measured patch latency and memory. I also appreciate the explicit acknowledgement that crop-centered scans are less suitable for some geometries and that the model is relatively heavy.

Critical weaknesses
  1. The methodological advance appears incremental, and baseline coverage is insufficient to establish its significance. Multi-scale feature alignment, refinement, and bidirectional fusion are established design patterns. The shared-weight direct/reduce-expand Mamba branches closely resemble the squeeze/reinforce branches in EM-Net, which the paper acknowledges. The main distinct contribution is therefore the specific radial scan and its combination with these components. The novelty argument should distinguish this contribution from prior spatial/anatomical modeling, including nnMamba and Topology-Guided Deformable Mamba (TGDM). TGDM uses costal-cartilage topology/centerline priors, whereas SPARNet uses a geometric crop center; they are related approaches rather than identical methods.

The experimental comparison should prioritize a well-configured nnU-Net v2, including an appropriate residual-encoder preset, which is absent despite nnU-Net being cited. Other relevant additions are nnMamba, EM-Net, and 3D TransUNet. The issue is incomplete coverage and configuration, rather than all baselines being outdated: SegMamba-V2 (2025) is already included. A common 500-epoch SGD recipe is useful for a controlled architecture study, but does not establish that each competitor is competitively configured. Please disclose baseline variants, adaptation/tuning choices, and resource budgets. This concern is supported by nnU-Net Revisited, which shows why configuration and compute-aware validation matter. For the complete architecture, Table 10 also increases capacity from 41.22M to 92.36M parameters; the plug-in controls support SLS, but do not fully isolate the benefit of the larger combined network.

  1. The evaluation is too narrow, with unresolved selection and reproducibility issues. Only two abdominal CT datasets are used: Synapse has 18 training and 12 validation volumes, and MSD Pancreas has 225 training and 56 validation volumes. Section B.2 selects the best checkpoint by validation Dice, while the reported evaluation appears to use these same validation splits; no separate untouched test set is described. This risks optimistic performance estimates, especially with scan variants and layer widths also investigated on Synapse. Repeated runs on one fixed split do not capture uncertainty from patient sampling; the number of runs/seeds is unspecified, and MSD results lack comparable uncertainty reporting. Broader generalization to different anatomy, modalities, or external cohorts remains unestablished.

There are also concrete inconsistencies that impede reproduction. Algorithm 1 does not explicitly initialize the Stochastic-In path and fuses path outputs before applying a single inverse permutation, whereas Section 5 says each path is inverse-permuted before fusion. Table 12 reports 85.34 Dice / 8.62 HD95 for uniform width 1.0, outperforming the chosen uniform width 3.0 (85.25 / 12.20) on both metrics, yet Section C.3 describes width 3.0 as offering the best Dice. These may be presentation errors, but the implemented coordinate alignment and model-selection rationale need to be resolved. Together with the limited novelty and incomplete baseline evidence, these issues lead me to 2: weak rejection.

Questions for authors

Clarifications that could materially affect my assessment:

  1. Were checkpoint selection, scan/width selection, and final evaluation performed on distinct patient sets? Please provide the exact split and selection procedure, the number of repeated runs and optimization/scan seeds, and explain which runs the tables summarize. If independent test, cross-validation, or patient-level uncertainty results already exist, please report them; otherwise qualify the generalization claims.

  2. Can you provide a comparison with properly configured nnU-Net v2 / nnU-Net ResEnc under disclosed, comparable resource budgets? Please identify the variants and configuration/tuning process for all baselines. Comparison with nnMamba and EM-Net would help establish progress within volumetric Mamba models; 3D TransUNet is another relevant comparator for global modeling and small-target segmentation. Please separate the controlled common-recipe comparison from results using each method's recommended configuration, if available, and clarify how capacity affects the full-network gains.

  3. Please reconcile Algorithm 1 with Section 5: where is the Stochastic-In path initialized, and are both path outputs restored to matched spatial coordinates before fusion? Please provide the corrected computation sequence corresponding to the reported implementation. Also explain the width-3.0 choice and revise the interpretation of Table 12, where width 1.0 has better Dice and HD95. Why do the HD95 changes in Table 11 not match subtraction from the baseline means in Table 2; do they use different runs or reference configurations?

  4. What is the precise methodological distinction from EM-Net's shared-Mamba branches and TGDM's topology-guided spatial modeling? Please compare the required priors, sequence construction, and coordinate restoration. A numerical TGDM comparison would be informative if its anatomy-specific components can be fairly adapted to the same tasks; if this is not appropriate, explain the incompatibility and provide a substantive related-work comparison.

Suggestions for future work, rather than demands for new experiments during rebuttal: evaluate additional anatomy/modalities or external cohorts, and quantify sensitivity to crop-center placement, window offsets, and scan-construction seeds. These would test the generality of the proposed geometric prior.

Reviewer_Un6oOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SPAR-Net combines scale-preserving multi-scale tokenization, complementary radial scanning, and bidirectional feature fusion for 3D medical image segmentation. Experiments on Synapse and MSD Pancreas report improvements in Dice and HD95, with additional analyses of scanning strategies and architectural components.

Critical strengths
  1. Meaningful empirical improvements. The final model improves Dice over SegMamba-V2 by 1.83 points on Synapse and 3.67 points on MSD Pancreas, including a 5.22-point tumor Dice gain. These gains accompany substantial HD95 improvements.
  2. Evidence beyond the full architecture. Scan-order comparisons and SLS integration into other Mamba backbones suggest that serialization can improve boundary accuracy independently of the complete SPAR-Net design.
Critical weaknesses
  1. The spatial-alignment procedure is inconsistent. Section 5 states that each path is inverse-permuted before fusion. Algorithm 1 instead fuses the paths in scanning space and applies a single inverse permutation afterward. Without additional alignment, the latter can combine features from different spatial locations. This may be a pseudocode error, but it leaves the core computation unclear.
  2. The specific benefit of radial serialization and its complementarity with SAFE are only partially isolated. Table 5 provides useful scan-order comparisons, including a two-path outward control, but does not report an explicitly matched bidirectional Cartesian baseline with the same path count, SSM evaluations, and parameter budget. Tables 5 and 6 study scan order and token scale separately, rather than directly testing their interaction with BPFB fixed. The existing results are encouraging, but do not fully isolate the claimed joint design benefit from generic bidirectional processing and multi-scale representation.
Questions for authors
  1. Are both paths spatially aligned before fusion in the implementation? Please reconcile Section 5 and Algorithm 1 with corrected pseudocode or indexing code.
  2. Does radial scanning outperform bidirectional Cartesian scanning with matched parameters, path count, and SSM evaluations? A SAFE-by-scan-order comparison with BPFB fixed would help establish complementarity.
  3. Could the authors compare against a recent 3D baseline, reporting Dice, HD95, parameters, and inference cost under comparable protocols?
  4. Were final evaluation cases separate from those used for checkpoint and configuration selection? Please clarify this separation and the number of repeated runs.
Reviewer_MWAROpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper presents a 3D Mamba segmentation network called SPAR-Net that jointly addresses multi-scale representation and sequence construction through scale-preserving tokenization and radial propagation, with bidirectional fusion between adjacent encoder levels. On Synapse dataset, SPAR-Net reports 85.25 Dice and 12.20 HD95. On MSD Pancreas SPAR-Net reports 64.02 mean Dice and 17.15 HD95.

Critical strengths
  1. The problem of how 3D medical volumes are serialized for SSMs is well motivated and interresting direction for medical imaging.

  2. The paper includes efficiency numbers and an honest discussion of failure cases (aorta, spleen trade-off).

  3. The HD95 reductions are large and consistent across organs, with organ-wise HD95 reported.

  4. SLS is tested as a plug-in on three Mamba backbones (Table 11).

Critical weaknesses
  1. One of the main issues in the paper is the baselines and protocol used by the authors. First, nnU-Net is cited but not compared. It is the standard reference and is typically stronger than the reported baselines on both datasets. In addition, all models use SGD with lr 1e-3 for 500 epochs, which likely underserves the transformers (i.e, Swin UNETR) 81.11 Dice is below commonly reported numbers. Were baseline hyperparameters tuned at all?

  2. Only two datasets are used, both abdominal CT, with small validation sets. There are no larger or more diverse benchmarks across different modalities (AMOS, BTCV-full, FLARE, BraTS/MRI).

  3. The idea of scan-order design for vision SSMs is an active area already: cross-scan in VMamba, bidirectional scans in Vim, local/zigzag scans, and spiral and space-filling-curve scans. None of these are discussed or compared. The dual-branch block is acknowledged as EM-Net-like. Pixel-shuffle multi-scale alignment is incremental.

  4. Table 11 deltas don't match Table 2 at all! The reported HD95 changes don't align with the baseline numbers. For U-Mamba, Table 2 gives 31.50 and Table 11 gives 30.16. For SegMamba, 30.69 to 19.82 is a drop of 10.87, not the claimed 10.46. For SegMamba-V2, 28.34 to 16.23 is a drop of 12.11, not the claimed 11.74.

  5. The comparison in Table 2 on Synapse has several missing strong and widely used 3D segmentation baselines. nnU-Net [1], and UNETR++ [2] and Slim UNETR [3] are not cited and not compared. Both are efficient hybrid transformers designed for the accuracy-efficiency trade-off in 3D segmentation. They are directly relevant to the paper's efficiency claims (Table 4), especially given SPAR-Net's 92.36M parameters and 1140 GFLOPs.

[1] F. Isensee, P. F. Jaeger, S. A. A. Kohl, J. Petersen, and K. H. Maier-Hein. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 2021. [2] A. Shaker, M. Maaz, H. Rasheed, S. Khan, M.-H. Yang, and F. S. Khan. UNETR++: Delving into Efficient and Accurate 3D Medical Image Segmentation. IEEE Transactions on Medical Imaging, 2024. [3] Y. Pang, J. Liang, T. Huang, H. Chen, Y. Li, D. Li, L. Huang, and Q. Wang. Slim UNETR: Scale Hybrid Transformers to Efficient 3D Medical Image Segmentation Under Limited Computational Resources. IEEE Transactions on Medical Imaging, 2024.

Questions for authors
  1. How does SPAR-Net compare to the mentioned methods and to baselines run using their recommended optimizers?

  2. Why was [3,3,3,3] chosen over [1,1,1,1]?

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

Related papers

Discussion (0)

Sign in to comment.