acceptodds
Withdrawn from ICLR 2027

ROLE-PRESERVING MULTISCALE EVIDENCE LEARNING FOR COMPACT MEDICAL IMAGE SEGMENTATION

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Compact medical image segmentation requires object-level discrimination without sacrificing low-contrast or irregular boundaries. Many compact encoder-decoder networks merge coarse semantics, regional structure, and fine texture into a single stream; an erroneous coarse response can therefore become a coherent false-positive object, whereas strong filtering can remove a genuine target. We propose a role-preserving multiscale evidence framework that assigns object evidence construction, regional continuity, boundary reconstruction, and class readout to explicit paths. At H/16, local and regional observations are related before a compact KAN mapper modulates a stable high-dimensional convolutional carrier. At higher resolutions, semantic, regional, and detail representations retain separate carriers and interact through geometric correspondence during transport. A cross-scale prototype head then allocates dense evidence to image-conditioned foreground and background prototypes instead of drawing an independent mask. Under a common three-seed protocol on three public benchmarks spanning ultrasound, endoscopy, and histopathology, our 5.79M-parameter model obtains mean IoUs of 73.97%, 85.06%, and 82.78% on BUSI, CVC, and GlaS, respectively. It achieves the highest BUSI and CVC mean IoU among the evaluated methods and improves the three-dataset average IoU from 79.52% for U-KAN to 80.60%. These results support role-explicit evidence transport as an effective architectural choice under the evaluated compact-segmentation protocol.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Medical Image Segmentation, Lightweight Neural Networks, Cross-Scale Representation Learning, Prototype Learning, Nonlinear Relation Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_Cu9iOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a role-preserving multiscale evidence learning framework for compact medical image segmentation. The central idea is to maintain distinct functional responsibilities for object evidence, regional continuity, boundary detail, and class allocation, rather than prematurely merging multiscale representations into a single feature stream. The proposed architecture incorporates local proposal and regional verification at low resolution, compact KAN-based relation mapping, role-specific cross-scale reconstruction, and an image-conditioned foreground/background prototype readout. These components aim to reduce cross-scale interference while preserving a relatively small parameter footprint. Experiments are conducted on three public medical image segmentation benchmarks, BUSI, CVC-ClinicDB, and GlaS. The proposed model achieves a three-dataset average IoU of 80.60% with 5.79M parameters. The authors also provide component ablations, role-fusion comparisons, and computational complexity analysis.

Critical strengths
  1. A coherent architectural motivation for multiscale feature interaction.

The paper addresses a relevant challenge in medical image segmentation: combining semantic information, regional structure, and fine-grained boundary details without introducing misleading responses. The role-preserving perspective provides a coherent organizational principle for the architecture. In particular, maintaining separate feature carriers while allowing relation-conditioned interactions is a potentially useful design strategy for compact segmentation networks.

  1. Evaluation across multiple imaging modalities with controlled ablations.

The experimental evaluation covers ultrasound, endoscopy, and histopathology, which exhibit substantially different appearance characteristics. The authors follow a common three-seed evaluation protocol and provide several component-level comparisons. The ablations involving KAN relation mapping, template conditioning, prototype allocation, and alternative fusion strategies offer useful evidence about the behavior of the proposed architecture.

Critical weaknesses
  1. Limited performance improvements relative to the architectural complexity and computational overhead.

Although the proposed architecture achieves the highest average IoU among the evaluated methods, the improvements over competitive baselines are relatively modest.

On BUSI, the method improves IoU over U-KAN by 1.03 percentage points. On CVC-ClinicDB, the improvement over Attention U-Net is only 0.34 points. On GlaS, the proposed method underperforms Attention U-Net by 0.55 points in IoU. The three-dataset average improvement over U-KAN is 1.08 points.

These relatively small gains should be considered alongside the additional computational cost. Despite having fewer parameters than U-KAN (5.79M versus 6.35M), the proposed method requires 7.19G MACs compared with 1.75G for U-KAN under the reported profiling protocol. Therefore, the reduction in parameter count does not translate into computational efficiency.

Furthermore, the reported standard deviations indicate noticeable variation on BUSI and GlaS, and the paper does not provide paired statistical comparisons establishing the reliability of the relatively small improvements.

These issues weaken the practical significance of the proposed design, particularly given its emphasis on compact medical image segmentation.

  1. The methodological novelty and generality of the role-preserving principle require stronger validation.

The proposed framework combines several established architectural operations, including residual convolutions, depthwise convolution, nonlinear relation mapping, gated feature modulation, learned cross-scale transport, and prototype-based classification.

Although their combination is carefully motivated, it remains unclear which aspects constitute a fundamentally new learning mechanism beyond a sophisticated arrangement of existing components.

The ablation results also reveal substantial dataset dependence. For example, the single-prototype variant outperforms the full model on CVC, while the parallel proposal-verification variant performs better on both CVC and GlaS. This suggests that the preferred configurations are not consistently optimal across different imaging modalities.

While the paper includes early-fusion, late-fusion, and no-role controls, it is unclear whether these alternatives are sufficiently matched in parameter count and computational cost to isolate the proposed information-organization principle from differences in model capacity.

In addition, the main comparison is restricted to a small set of established segmentation architectures. More recent compact segmentation methods discussed in the related work are not included in the quantitative evaluation. This limits the evidence for the competitiveness of the proposed approach relative to current efficient segmentation designs.

Furthermore, the architecture figures appear to be programmatically generated SVG diagrams, with several mathematical symbols and expressions displayed as raw, unrendered text rather than properly typeset notation, which further undermines the presentation quality and professionalism of the manuscript.

Questions for authors

Q1. Computational efficiency and practical trade-offs.

The proposed model has fewer parameters than U-KAN but requires approximately four times as many MACs, while improving the three-dataset average IoU by only 1.08 points. Could the authors provide actual inference latency, peak GPU memory usage, and throughput comparisons under identical hardware conditions? These measurements would help clarify the practical value of the proposed architecture as a compact segmentation framework.

Q2. Isolating the contribution of role preservation.

Could the authors clarify whether the early-fusion, late-fusion, and no-role variants are matched to the full model in parameter count and computational complexity? If not, could they provide capacity-matched comparisons to establish that the improvements arise from the role-preserving information flow rather than additional feature-processing capacity?

Q3. Robustness and dataset-dependent component behavior.

The performance differences between the full model and several variants are relatively small, and some alternative configurations outperform the complete architecture on individual datasets. Could the authors provide paired statistical comparisons across seeds and further explain why the single-prototype and parallel proposal-verification variants outperform the full model on certain datasets? This would clarify the robustness and generality of the proposed design choices.

Q4. Evaluation protocol and baseline coverage.

Could the authors clarify whether the BUSI and CVC data partitions account for patient or acquisition-sequence grouping, where applicable? Additionally, could they explain why more recent compact segmentation methods, including representative architectures discussed in the related work, are not included in the quantitative comparisons? Stronger contemporary baselines would help establish the significance of the reported improvements.

Reviewer_6Z5yOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a 5.79M-parameter U-shaped segmentation network organized around what it calls a role-preserving constraint: object evidence, regional continuity, boundary detail and class allocation keep separate carriers and may exchange only conditioners, never a fused replacement. At H/16 a local proposal and a regional verification are combined through agreement/conflict terms and a compact KAN mapper that predicts an affine field for a 128-channel convolutional carrier; at H/8-H/1 three carriers share one learned geometry field while keeping independent value projections; and the readout matches pixel evidence to image-conditioned foreground/background prototypes, letting context adjust mixture weights rather than emit its own mask. Under a common three-seed, six-baseline protocol, it reports macro IoU 80.60% versus 79.52% for U-KAN, best BUSI and CVC IoU, and a GlaS loss to Attention U-Net.

Critical strengths
  1. The paper tests its own organizing principle rather than only its parts: the right panel of Table 3 compares full role separation against early fusion, late fusion and no role constraint, and the per-dataset breakdown is reported instead of being hidden by the average. Collapsing the roles costs 3.3-3.8 IoU on BUSI and 2.0-3.0 on GlaS while leaving CVC unchanged, which is the right experiment for a claim that is architectural rather than component-level.

  2. The KAN-versus-MLP choice is settled by an approximately parameter-matched replacement inside the same architecture (A1), and the paper reports where the design loses — GlaS IoU/F1 to Attention U-Net, CVC HD95 to the convolutional baselines — and states that benchmark-informed development makes these fixed-benchmark numbers rather than external-cohort estimates. That disclosure is more than most submissions offer and it is what allowed me to audit the tables quickly.

Critical weaknesses
  1. The figures do not describe the method they are supposed to illustrate. Figure 1's legend assigns "stride=1/2/3" to 3x3 convolutions, including a stride-3 convolution that never appears in Section 3.2 where downsampling is done by pooling; ReLU and GeLU are drawn as the same decorative glyph; several labels render as ghosted duplicates (Figure 2) or spill outside their boxes (Figure 4); and Figure 2's central panel shows a decorative blob where the "proposal-conditioned feature" should be described. With a 3D-gradient/icon vocabulary that is unlike the rest of the typesetting, this reads as image-model output rather than a schematic, and it leaves the architecture unreconstructable from the figures.

  2. The contribution is a recombination, and the evidence does not establish even a modest gain. Two-stream region/boundary separation, KAN on compressed relations, prototype allocation and content-aware upsampling all pre-exist in the cited literature (BiSeNet, Gated-SCNN, U-KAN, OCR/ProtoSeg, CARAFE/SAPA), and Eq. 1 is a generic relation-conditioned residual that any multi-branch U-Net satisfies, so "role preservation" predicts nothing beyond fusion order. The 1.08-point macro-IoU gain over U-KAN is about 1.8 standard errors with no significance test, and the five baselines exclude the compact SOTA the paper itself cites (EMCAD, HCI, Rolling-UNet, EGE-UNet).

Questions for authors

Q1. How were Figures 1-4 produced? Please confirm the generator or drawing tool, explain the stride-2/stride-3 convolution entries in Figure 1's legend that do not appear in Section 3.2, and redraw them as schematics that match the implemented model.

Q2. With three seeds, the 1.08-point macro-IoU gain corresponds to roughly 1.8 standard errors. Can you report a paired test across seeds, or confidence intervals on the macro average, for Table 2 and Table 3?

Q3. Why are EMCAD, HCI, Rolling-UNet and EGE-UNet absent from Table 2 when Section 2 places this work in that compact line? One compact SOTA baseline retrained under your protocol would decide whether the claim extends beyond the five evaluated methods.

Q4. A1 (MLP) beats the full model on CVC, and A7 (single prototype pair) beats it by 1.43 IoU on CVC while losing 4.50 on BUSI. How is the three-mode readout justified if a single prototype pair is better on one of three benchmarks?

Q5. What are the parameter and MAC counts of each ablation variant, not only of the full model? For A1 the substance is the matched-capacity comparison, so please give the exact parameter gap between the KAN mapper and the MLP replacement.

Reviewer_HfRnOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a compact U-shaped segmentation network that assigns object verification, regional continuity, boundary detail, and class allocation to distinct paths. It combines local proposal and regional verification at H/16, compressed KAN relation mappings, role-specific cross-scale transport, and image-conditioned foreground/background prototypes. Under a common three-seed protocol, the 5.79M-parameter model obtains IoUs of 73.97%, 85.06%, and 82.78% on BUSI, CVC-ClinicDB, and GlaS, respectively.

Critical strengths
  1. Clear architectural motivation. Separating object, regional, and detail responsibilities is more interesting than adding another undifferentiated fusion block. The paper explains how premature mixing could amplify false positives or suppress weak targets.
  2. Sensibly scoped KAN usage. Applying KAN to compressed relation representations while retaining a convolutional carrier is a defensible design. The approximately parameter-matched MLP ablation directly examines this choice.
  3. Useful experimental transparency. The paper evaluates three imaging domains, reports three-seed means and standard deviations, includes component and role-fusion controls, and provides individual final-model runs. It also acknowledges nonuniform gains and the limits of benchmark-based development.
Critical weaknesses
  1. Limited novelty and insufficient validation of the central concept. The system combines familiar residual conditioning, multi-scale feature processing, content-aware transport, and prototype readout. The main claimed contribution is their role-preserving organization. However, the operational definition in Eq. (1) is broad enough to describe many gated residual architectures, and retaining a carrier does not itself demonstrate that it learns the intended object, regional, or boundary responsibility. The existing fusion controls are useful, but there are no role-specific probes, targeted perturbations, or quantitative measures of cross-role interference.

  2. The controls do not isolate role separation from capacity and operators. Table 3 names early fusion, late fusion, and no-role variants, but the current version of paper does not specify their exact replacements, widths, parameter counts, or MACs. Calling them matched controls in the conclusion does not establish which quantities are matched. There is also no reported baseline using the same compact residual encoder with a conventional decoder. Comparisons with a 34.53M-parameter U-Net therefore cannot determine how much of the parameter reduction comes from the narrower backbone. A fully specified comparison with the same encoder, similar compute, and conventional fusion is needed to attribute the improvement to role preservation.

  3. The practical benefit is small relative to the computational burden. Table 4 reports 7.19G MACs, versus 1.75G for U-KAN and 0.57G for UNeXt: approximately 4.11 and 12.61 times as many operations. Relative to U-KAN, the parameter reduction is only 8.8%, while macro IoU improves by 1.08 points. Section 4.4 correctly limits its claim to parameter compactness; nevertheless, there are no latency, throughput, peak-memory, FPS, or training-cost measurements to establish a practical advantage.

  4. The comparison set is too limited for a compact-segmentation contribution. Table 2 includes only U-Net, Attention U-Net, UNet++, UNeXt, and U-KAN. EMCAD “https://arxiv.org/abs/2405.06880” (Rahman et al., CVPR 2024) is already cited but not evaluated, despite being directly relevant to efficient multi-scale decoding. There is no transformer or state-space baseline, such as VM-UNetV2 “https://arxiv.org/pdf/2403.09157” (Zhang et al., 2024), the latter also studying polyp segmentation.

  5. High variance and marginal gains vs. simple baselines. As shown in Table 2, on the BUSI dataset, the proposed method achieves 73.97 +/- 1.46 IoU, which is only ~1 point higher than the much simpler U-KAN baseline (72.94 +/- 0.69). Not only is the variance of the proposed model more than double that of U-KAN, but the performance on GlaS is virtually identical (82.78 +/- 1.19 vs 82.60 +/- 0.96). Given the sheer number of complex, hand-crafted modules introduced (proposal-verification, role-specific transport, multi-prototype readout, KAN relations), a ~1 point average improvement with higher variance strongly suggests that the gains may simply be due to seed noise rather than the architectural novelty. The paper lacks statistical significance tests (e.g., paired t-tests, confidence intervals) to prove otherwise.

  6. Potential cherry-picking and contradiction between qualitative and quantitative results. In Table 2, Attention-UNet actually achieves better quantitative results on the GlaS dataset (IoU 83.33 +/- 1.34 and F1 90.76 +/- 0.85) compared to the proposed method (IoU 82.78 +/- 1.19 and F1 90.49 +/- 0.70). However, in the prediction figure (e.g., Figure 5), the Attention-UNet results are shown to be visually much worse than the proposed method. This raises concerns about whether the qualitative examples are cherry-picked and presents a contradiction between the visual evidence and the overall quantitative results.

  7. Ablations leave important design choices unresolved. The existing component ablations should be credited, but they do not separately isolate shared geometry versus independent geometry, bounded versus unbounded conditioning, the two-slot anchor, or prototype allocation versus a conventional segmentation head. Some ablation results are compatible with cross-dataset trade-offs, but do not show that the full collection of mechanisms is necessary. There is also no sensitivity study for the fixed numerical bounds.

  8. Architecture figures are difficult to audit against the equations. Figures 1, 3, and 7 rely on many small operator glyphs and abbreviated labels, while the mapping from graphical paths to named carriers, residual updates. Figure 4 compresses prototype construction, overlap removal, and mixture allocation into a schematic without exposing their distinct operations. The caption of Figure 6 also uses the unexplained label E3-A.

Questions for authors

See Weaknesses,

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

Related papers

Discussion (0)

Sign in to comment.