List of Papers Browse by Subject Areas Author List
Abstract
Accurate 3D segmentation is central to quantitative lesion assessment and anatomy mapping for clinical planning and follow-up. Thin, elongated, and fine anatomical/pathological structures (e.g., ves sels) are a particularly stress case: small boundary error can disrupt apparent continuity and bias quantitative measurements. In encoder–decoder networks (e.g., U-Net), repeated downsampling and fixed-grid convolution blur/alias fine structures and weaken orientation cues, so early mistakes propagate across scales. We propose a geometry-guided local operator that steers where features are sampled (rather than deforming convolutional kernels) under a single formulation for both feature refinement (stride 1) and resolution reduction (stride > 1). At each voxel, it predicts a local orientation and bounded step sizes, samples symmetrically along these directions, and transforms paired samples into compact geometric/boundary cues with lightweight mixing; a cross-scale consensus aligns encoder and decoder features at skip connections to reduce geometric mismatch. Replacing all stride 1 and stride 2 operators in a 3D U-Net yields consistent improvements on BraTS, MSD Hepatic Vessel, and TDSC-ABUS, with notably better boundary metrics (e.g., BraTS Dice 86.1→88.9, HD95 7.1→6.2; TDSC-ABUS HD95 39.1→27.8) while substantially reducing parameters (2.3M→0.8M). We further demonstrate that the operator plugs into other backbones (e.g. nnU-Net, Swin-UNETR, and MedNeXt) without changing their macro-architectures while providing consistent performance gains. Code available at GeoSample repo.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2940_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/AndyWanng/GeoSample
Link to the Dataset(s)
HepaticVessel dataset: http://medicaldecathlon.com
TDSC-ABUS dataset: http://tdsc-abus2023.grand-challenge.org
BraTS2023 dataset: https://www.synapse.org/Synapse:syn51156910/wiki/621282
BibTex
@InProceedings{WanSiz_Steer_MICCAI2026,
author = { Wang, Sizhe AND Peiris, Himashi AND Chen, Zhaolin},
title = { { Steer the Sampling, Not the Kernel Grid: Geometry-Guided Sampling Operator for Volumetric Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes GeoSample, a geometry-guided local operator for 3D medical image segmentation that replaces fixed-grid convolution/pooling with voxel-wise, rotation-constrained symmetric sampling, unifies refinement and downsampling in one formulation, and adds a Consensus Field for geometry-aware skip alignment across scales.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper proposes a novel geometry-guided local operator, GeoSample, that directly targets a core weakness of volumetric segmentation: fixed-grid convolution and pooling often blur thin, elongated, and anisotropic structures. Instead of redesigning the whole backbone, the method rethinks the sampling primitive itself by predicting a voxel-wise local 3D rotation frame and adaptive bounded step sizes, then steering structured symmetric sampling along those directions. This is interesting because it introduces explicit local geometric structure into feature extraction in a way that is better matched to boundary-sensitive medical targets.
2.A major strength is the elegant formulation of geometry-aware cues from the sampled features. The paired symmetric samples are converted into compact average, gradient-like, and curvature-like signals using step-size-normalized finite differences. This is more than a simple deformable sampling variant: the formulation has a clear geometric interpretation, explicitly separates even/context and odd/directional-change information, and is designed to make the representation more stable across different local scales.
3.The method is unified and architecturally coherent. GeoSample is not limited to one component of the network, but is formulated as a single operator that can replace both stride-1 feature refinement and stride>1 downsampling. This is a strong design choice because the same geometric principle is applied consistently across scales, rather than only improving one isolated stage of the encoder-decoder pipeline.
4.The paper includes an additional technically meaningful contribution through the rotation-consistent Consensus Field. This module fuses geometric fields from encoder and decoder features at skip connections using quaternion-based rotation fusion and adaptive gating. This is an original and useful idea because it addresses cross-scale geometric mismatch at the level of the local sampling geometry itself, rather than only fusing appearance features.
5.The experimental evaluation is reasonably broad and supports the main motivation of the work. The method is tested on three public 3D datasets spanning three modalities, namely BraTS (MRI), MSD Hepatic Vessel (CT), and TDSC-ABUS (ultrasound). This diversity is a strength because it demonstrates that the method is not tailored to a single imaging setting and appears to generalize across substantially different anatomical targets and image characteristics.
6.The reported empirical gains are particularly convincing on boundary-sensitive metrics, which is well aligned with the claimed motivation. In the operator-level comparison, the proposed method achieves the best overall mean Dice/HD95/ASSD, performs best on BraTS, achieves the best Dice and HD95 on Hepatic Vessel with near-best ASSD, and substantially improves HD95 and ASSD on TDSC-ABUS while keeping Dice competitive. These results are important because they suggest that the operator is especially effective for preserving fine boundary information rather than only improving coarse region overlap.
7.The plug-in validation across multiple backbones increases the practical value of the paper. The authors show that the operator can be integrated into nnU-Net, Swin-UNETR, and MedNeXt without changing their macro-architectures, and that it yields consistent improvements in overlap and boundary metrics. This strengthens the contribution because it suggests the method is a reusable operator-level improvement rather than a one-off architecture-specific trick.
8.Another strong aspect is the efficiency-accuracy tradeoff. In the U-Net setting, the proposed method reduces parameters from 2.3M to 0.8M and FLOPs from 194.8G to 108.9G while still improving segmentation quality. This is noteworthy because geometry-aware or adaptive operators often increase model cost, whereas here the method appears to deliver both better accuracy and lower complexity, which improves its practical attractiveness for real 3D medical segmentation systems.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The methodological novelty appears narrower than the paper’s positioning suggests. Many core ingredients are adjacent to existing directions already discussed in the manuscript, including deformable convolutions (Dai et al. , ICCV 2017; Zhu et al. , CVPR 2019), volumetric deformable variants such as DeformUX-Net (Lee et al. , 2023), sparse deformable sampling in OBELISK-Net (Heinrich et al. , MedIA 2019), spatial transformer based geometric warping (Jaderberg et al. , NeurIPS 2015), and equivariant/steerable 3D models such as Group Equivariant CNNs (Cohen and Welling, ICML 2016), 3D Steerable CNNs (Weiler et al. , NeurIPS 2018), and SE(3)-Transformers (Fuchs et al. , NeurIPS 2020). The main distinction here is the use of a rotation-constrained local frame with symmetric paired sampling and differential-token construction, which is interesting, but the manuscript does not yet make a sufficiently sharp case for why this goes substantially beyond a structured combination of prior ideas.
2.The experimental validation does not fully support the strongest claims about thin-structure preservation, topology, and clinical relevance. The paper repeatedly motivates the method by arguing that small boundary errors can disconnect branches or distort clinically meaningful topology, especially for thin structures such as vessels. However, the evaluation mainly reports Dice, HD95, and ASSD. These metrics are useful, but they do not directly quantify connectivity or topology preservation. Since the paper itself cites topology-preserving segmentation work (Hu et al. , NeurIPS 2019), the absence of topology-aware evaluation is noticeable. For the hepatic vessel task in particular, metrics such as clDice, centerline Dice, connected-component errors, branch recall, or topology error counts would be much more aligned with the paper’s core claims.
3.The comparison is not fully fair from a model-capacity perspective, which weakens the “better and more efficient” conclusion. In the U-Net setting, the proposed model is substantially lighter than the compared baselines (0.8M parameters and 108.9G FLOPs versus 2.3M/194.8G for baseline U-Net, 2.5M/203.3G for DCNv1, 2.6M/209.3G for DCNv2, and 2.7M/384.5G for Dynamic Downsampling). As a result, the current comparison is not capacity-matched. Some of the performance difference may come from a different parameterization or regularization regime rather than purely from the proposed operator itself. Stronger evidence would require parameter-matched or FLOP-matched baselines, or at least a lighter baseline U-Net with similar capacity.
4.The empirical gains are promising but not yet statistically convincing. All methods are evaluated using a fixed-seed 75/10/15 split, and the paper reports only single summary numbers without repeated runs, standard deviations, confidence intervals, or significance testing. This is important because some gains are modest or mixed. For example, on MSD Hepatic Vessel, the proposed method achieves the best Dice and HD95 but not the best ASSD; on TDSC-ABUS, the Dice improvement over the strongest baseline is small and the main gain is in boundary metrics. These results may still be meaningful, but without statistical analysis it is difficult to judge their robustness. For a TMI submission, multi-seed evaluation or significance testing would make the conclusions much more convincing.
5.The plug-in generalization claim is only partially supported. The paper argues that the operator can be plugged into multiple backbones, but the validation is limited to a single dataset, TDSC-ABUS. Moreover, the intervention is not consistent across architectures: in nnU-Net the authors replace all stride-1 and stride>1 blocks, whereas in Swin-UNETR and MedNeXt they replace only downsampling blocks to preserve backbone-specific refinement. Because the integration strategy differs across backbones, the evidence for general backbone-agnostic applicability is weaker than the paper suggests. The results are encouraging, but they do not yet fully establish generality.
6.The ablation study is too limited for a method with several nontrivial design choices. The current ablation only removes the differential terms and the Consensus Field. However, the proposed method also relies on several specific choices, including SO(3) frame prediction, bounded step sizes, symmetric paired sampling, step-size normalization, odd-signal preservation during pooling, and quaternion-based fusion at skip connections. The manuscript does not isolate the contribution of these components. It would be particularly useful to test symmetric versus unconstrained sampling, bounded step sizes versus free offsets, SO(3) frame prediction versus direct vector prediction, and the effect of K and the normalization by rk. Without these experiments, it remains unclear which parts are truly essential.
7.Reproducibility and practical efficiency analysis remain insufficient. Although the paper describes the main operator design and states that the same training pipeline is used across methods, many implementation details remain under-specified, including the loss function, optimizer, learning-rate schedule, training duration, data augmentation, and exact hyperparameter settings such as rmin, rmax, and epsilon. In addition, the conclusion explicitly acknowledges remaining memory and wall-clock overhead from interpolation-based sampling, as well as early-training stability issues. This makes it important to report actual runtime and GPU memory usage, not only FLOPs. At present, the efficiency discussion is incomplete from a practical deployment perspective.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses an important problem in 3D medical image segmentation and proposes a technically interesting geometry-guided sampling operator for both refinement and downsampling. I find the idea meaningful and the multi-dataset results generally promising, especially on boundary-related metrics.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
After reading the rebuttal, I am satisfied that the authors have adequately addressed my main concerns, and I support acceptance. In particular, the rebuttal clarifies the intended novelty of GeoSample relative to deformable convolution, OBELISK-style sampling, and related geometric methods, and frames the contribution more appropriately as a geometry-aware local sampling primitive for refinement and downsampling rather than a claim of broad topological modeling. The authors also gave a clearer explanation of the compact U-Net scaffold and the reported parameter counts, which resolves my earlier concern about comparison settings. In addition, the integration strategy for nnU-Net, Swin-UNETR, and MedNeXt is now much better justified, especially the decision to preserve backbone-specific refinement blocks in transformer/ConvNeXt-style architectures while replacing the common downsampling and skip-alignment components. Although some limitations remain, such as the lack of topology-specific evaluation and more extensive ablations, I think these points no longer undermine the main contribution. Overall, the paper presents a meaningful and technically coherent operator-level contribution, and the rebuttal makes the claims more precise and the experimental design more understandable, which is sufficient for me to maintain an accept recommendation.
Review #2
- Please describe the contribution of the paper
The paper proposes a geometry-guided sampling operator, GeoSample, for 3D segmentation. Instead of modifying convolution kernels, its core idea is to predict local orientations and step sizes, perform symmetric sampling along these directions, and then construct gradient-like and curvature-like cues, while unifying stride-1 refinement and stride-2 downsampling within a single framework. In addition, the paper introduces a Consensus Field at skip connections to achieve geometric alignment between encoder and decoder features. The motivation of the method is fairly well developed, the formulation is relatively clear, and the experiments cover three datasets, namely BraTS, MSD Hepatic Vessel, and TDSC-ABUS, together with plug-in validation on different backbones. Overall, this is not merely an engineering-heavy combination of existing components, but rather a reasonably unified operator design.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1）The paper presents a relatively unified operator design rather than a simple combination of engineering components. 2）A notable strength is that the proposed method integrates both feature refinement and downsampling within a single framework. 3）The paper addresses a meaningful problem, namely the loss of boundary and geometric information caused by fixed-grid sampling in 3D segmentation. 4）The experimental results are consistently positive across MRI, CT, and ultrasound, suggesting good cross-modality applicability.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1）In all experiments, the number of sampling directions K is fixed to 3, corresponding to a rotated 3D coordinate frame. A potential concern is whether only three symmetric pairs of samples, namely six sampled points in total, are sufficient to capture local topology for complex vascular branching patterns or extremely thin structures. 2）The ablation study shows that removing the Consensus Field leads to a substantial performance drop, with the Dice score on TDSC-ABUS decreasing from 67.4 to 58.4.This raises a concern that the proposed operator may be highly sensitive to cross-scale alignment, making its effectiveness strongly dependent on this additional module. 3）For the central claim regarding thin and elongated structures, the inclusion of the Hepatic Vessel dataset is helpful, but the validation is still not sufficiently thorough. For example, the paper does not clearly demonstrate whether the method truly preserves topology, connectivity, or fine-branch recall, rather than merely improving boundary distance metrics. 4）Recent work on deformable convolution remains active, especially in CVPR/MICCAI/TMI. However, authors are recommended to include more recent papers for a more comprehensive comparison, such as DCNv4, DefMamba, etc. 5）As far as I am concerned, MedNext with the base version has the parameter size of about 10M. Thus, I am not sure whether the “3M” value is correct enough. Please have a check. Also, why does the proposed operator bring a parameter decrease on nn-UNet, but an increase on MedNeXt and Swin UNETR? Besides, since the core component of Swin UNETR is self-attention block, how is the proposed operator incorporated to Swin UNETR, by replacing the convolution kernels or self-attention blocks? Please give a more detailed description.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My overall score is mainly based on the following considerations. The paper addresses an important problem in volumetric medical image segmentation and presents a reasonably coherent operator design, with clear motivation and generally positive empirical results across multiple datasets. However, several factors prevent me from giving a more positive recommendation. Please see the weakness part. I will consider to increase my score if these concerns are well addressed.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The rebuttal has addressed part of my concerns, and I appreciate the additional clarifications from the authors. However, I believe the manuscript still requires substantial revisions before the final version, including the justification and analysis of the fixed sampling strategy, comparisons with more recent deformable convolution methods, more thorough validation for topology and connectivity preservation of thin structures. Although the other two reviewers are inclined to accept this paper, I still think there is a distance before publication to MICCAI.
Review #3
- Please describe the contribution of the paper
The authors propose a set of replacement operators for standard convolutions and strided convolutions in pooling operations rooted in geometry aware sampling. Their method essentially consists of a sampling operation which adapts to the image and is defined by a per voxel geometry field and a differentiable mixing for the actual feature learning. There is also a consensus field for merging encoder to decoder representations.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The authors present a novel operator in line with that of deformable convolutions. They infact do make this comparison in their ablations, thereby highlighting the throughline between the two methods. Their operator seems to be a plug-in for existing architectures and is demonstrated as such on 3 representative deep networks and multiple datasets.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.While the nature of the methodological contribution may necessitate mathematical density, I believe it is difficult to understand in places. The authors should consider the readability of a diverse audience for their methods section in particular. It is overframed in mathematical language and would not necessarily help a reader easily reconstruct the forward pass from their description.
2.While it is understandable that MICCAI is a venue with limited page length, the authors should definitely consider that the Figure 1 is rather small and unreadable.
3.While this is not really a network proposal, contextualizing the results of a strong network (ResEncL, MedNeXt-L) after applying the proposed operator, against popular baseline networks would have been useful to gauge the benefits of the operator in state-of-the-art benchmarking.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The authors propose a plausible and useful operator in lieu of convolutions with plug and play capabilities for existing architectures. They compare against their nearest neighbours in the form of deformable convolutions and describe their method in detail. However, their manuscript does limit readability by mathematically overframing their methodology.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I appreciate the authors’ attempt to address my review and maintain my positive score from the initial review.
Author Feedback
We thank the AC/reviewers. Related comments are grouped below. [AC1/R1C1/R2C4] Novelty. GeoSample is fundamentally different from deformable aggregation: we sample to measure voxel-wise geometry, while DCN/OBELISK-style methods sample shifted locations to aggregate features. They learn offsets/patterns for mixing; we learn a local SO(3) frame and bounded steps, then encode context, directional change, and curvature-like variation. STN/equivariant methods warp features or enforce transformation rules, whereas GeoSample uses geometry to construct voxel representations before refinement/downsampling. DeformUX changes blocks, recent DCNs refine offset aggregation, and DefMamba focuses on scan/mixing. Thus they share ingredients but not the same 3D refinement/downsampling primitive; DCNv1/v2 are the closest controlled baselines. We will clarify this in Introduction. [AC1] Performance. We agree the Dice margin is modest, so we do not claim a large Dice-only gain. GeoSample improves HD95 25.6→22.8 and ASSD 4.94→4.72, with parameters/FLOPs 2.5M/203.3G→0.8M/108.9G. This matches the goal: compact geometry-aware sampling for stable boundaries. [R2C1/R2C2/R1C2/R1C6/R2C3] K parameter/topology insights/ablation. K=3 is sufficient for the intended local measurement. It forms a complete 3D frame; we do not claim one six-point stencil recovers whole-structure topology. Each voxel measures context, directional change, and curvature-like change along this frame, while branches are represented by dense measurements across layers/scales. Consensus Field is part of the operator, not an external add-on; our ablation targets the two highest-level design choices—local geometry measurement and cross-scale consistency—showing both are needed for reliable skip connections. Topology motivates why boundary errors matter; our evidence supports boundary/local-geometry quality. Topology-specific evaluation is complementary. We will add this to Discussion. [AC2/R1C5/R2C5] Plug-in and integration. Integration is functionally consistent: common insertion points are resolution changes and skip connections. In nnU-Net, stride-1 refinement and stride>1 reduction are standard convolutional operators, so replacing both is direct substitution. In Swin-UNETR/MedNeXt, stride-1 refinement is the defining self-attention or ConvNeXt-style block; replacing it would change the backbone family rather than test plug-in compatibility. We therefore preserve backbone-specific refinement and replace only downsampling, with Consensus Field at skips. Swin-UNETR self-attention blocks are not replaced. For MedNeXt, we used the base block from official repo with base channel 16, yielding 2.7M params. [AC2/R1C3/R1C7] U-Net. The 2.3M U-Net is a compact fixed scaffold. It has four levels with base channel 16.Each encoder level uses one stride-1 3x3x3 refinement operator and one stride-2 3x3x3 reduction operator; the decoder mirrors this layout. All variants replace the same locations under the same configuration. Thus, the comparison controls scaffold, width, split, and training, rather than equalising parameter count. The lower cost reflects structural efficiency: GeoSample improves segmentation quality with fewer learned weights. [R1C4/R1C5/R5C3] Experimental scope. We agree repeated runs, multi-dataset plug-ins, and larger backbones would be interesting. However, our goal was to answer two operator-level questions: (i) whether GeoSample transfers across modalities/anatomies under the same U-Net scaffold, and (ii) whether it can be inserted into other SOTA pipeline/backbones with consistent gains. Future work could include testing its scalability with ResEncL/MedNeXt-L and other larger models. We will include this in Discussion. [R5C1/R5C2/R1C7] Clarity/reproducibility. We will add a forward-pass summary, enlarge Fig. 1, and add missing training details. The submission included an anonymous Git repo with code/reproducible settings; we will make it public upon acceptance.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The reviewers found that the proposed sampling operator is novel and addresses a clinically relevant problem. Experiments were conducted on datasets covering different imaging modalities and anatomical structures.
However, the reviewers noted that the paper is difficult to follow due to the number of formulations. In addition, several important experimental details are missing.
The authors should clarify (but not limited to) the following points in the rebuttal:
1.What are the key advantages of the proposed sampling operator compared to existing methods, such as Deformable Conv and other recent approaches? Notably, the performance of Deformable Conv v1 is very similar to that of the proposed framework, with the differences in Dice scores less than 1%.
2.More details should be provided for the baseline U-Net, especially the number of conv layers. The reported parameter count of 2.3M appears unusually small for a 3D U-Net model. In addition, more implementation details for MedNeXt should be provided, as pointed out by Reviewer #2. Note that new results, experiments, or data beyond what was provided at the paper’s original submission are not permitted in the rebuttal.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Two out of the three reviewers are satisfied with the rebuttal, and I share their assessment. Therefore, I recommend acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The paper has a methodology ambiguity, though 2 reviewers have positive ratings. For example, after reading, it’s still unclear to me what the geometry signal is that is used for sampling. Many jargon statements and claims make the paper hard to understand.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper proposes GeoSample, a geometry-guided local operator for 3D medical image segmentation that replaces fixed-grid convolution and pooling with voxel-wise, rotation-constrained symmetric sampling. Experiments on BraTS, MSD Hepatic Vessel, and TDSC-ABUS show consistent improvements, particularly on boundary metrics, with reduced parameter count (2.3M→0.8M). Reviewers agreed that the operator-level rethinking of sampling is a meaningful and technically interesting contribution, the elegant formulation separates even/context and odd/directional-change information, the unified design across stride-1 and stride>1 is architecturally coherent, the Consensus Field is an original geometry-aware skip-fusion idea, the evaluation across three modalities is broad, the plug-in validation across nnU-Net/Swin-UNETR/MedNeXt is practically useful, and the efficiency-accuracy tradeoff is favorable. They raised several concerns: (1) novelty positioning relative to deformable convolutions, OBELISK-Net, STN, and equivariant models needs sharpening; (2) topology-aware evaluation (clDice, branch recall, connected-component errors) is missing despite the topology motivation; (3) the capacity-matched baseline comparison is not fully isolated, since the proposed model is lighter than baselines; (4) single fixed-seed evaluation without significance testing; (5) plug-in generalization is only on TDSC-ABUS, with different integration strategies across backbones; (6) the ablation does not isolate several non-trivial design choices (SO(3) frame, K, symmetric sampling, normalization); (7) the methods section is mathematically dense and Figure 1 is hard to read; and (8) reproducibility details (hyperparameters, training settings) are under-specified.After rebuttal, the authors clarified GeoSample’s distinction from DCN/OBELISK/STN-style methods (measuring geometry vs. aggregating shifted features), justified the compact U-Net scaffold, explained the backbone-specific integration choices, and committed to improving figure clarity and reproducibility details. Both reviewers maintain or raise their recommendations to Accept. The AC leans toward acceptance.