acceptodds
Under review as a conference paper at ICLR 2027

GPS: Geometry-aware and Peripheral-background Synergy for Cross-Domain Few-Shot Medical Image Segmentation

Abstract

Few-shot medical image segmentation (FSMIS) aims to generalize to novel classes with minimal annotations, yet its effectiveness is severely hindered by the challenge of domain shift, where substantial inter-domain discrepancies (e.g., MRI to CT) often lead to significant performance degradation. Current cross-domain FSMIS methods primarily rely on volatile appearance-based features or computationally expensive graph-matching, failing to capture stable anatomical representations. In this paper, we observe a crucial yet under-explored phenomenon: while domain-specific textures vary drastically, the geometric topology of organs and their surrounding peripheral-background distributions remain remarkably consistent across divergent domains. Based on this insight, we propose a novel framework termed Geometry-aware and Peripheral-background Synergy (GPS) for cross-domain FSMIS. GPS leverages these stable invariants as "Anatomical Anchors" to steer the model beyond domain-sensitive biases. Specifically, GPS consists of three synergistic modules: (1) a Geometry-Aware Injector (GAI) that modulates feature responses based on geometric shape priors to distill domain-invariant representations; (2) a Peripheral-aware Prototype Mining (PPM) module that constructs multi-prototypes for both the target organ and its surrounding tissues; and (3) a Prototype Synergy Refinement (PSR) module that exploits foreground-background complementarity for fine-grained segmentation. Extensive experiments on three popular medical image benchmarks demonstrate that our GPS significantly outperforms state-of-the-art methods, exhibiting superior robustness and generalization across diverse domains.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
medical image segmentation, few-shot learning, cross-domain

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_my8vOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes GPS for cross-domain few-shot medical image segmentation (CD-FSMIS). Motivated by the observation that organ geometry and peripheral-background distributions stay stable across domains, it treats them as “anatomical anchors”. Three modules follow: GAI, which modulates feature channels via a dual-ring radial contour descriptor of the support mask; PPM, which mines multi-prototypes for the foreground and for the peripheral region inside the bounding box; and PSR, which fuses both similarity maps for coarse-to-fine refinement. On abdominal CT/MR, cardiac LGE/b-SSFP, cross-institution prostate, and two extreme-shift tasks, GPS achieves the best mean Dice under 1-way 1-shot at lower cost than C-Graph.

Critical strengths

1.Comprehensive and controlled evaluation: seven public datasets, nine baselines retrained under a shared backbone and identical data splits, plus efficiency and hyperparameter-sensitivity analyses.

2.Clear figures and a readable narrative: Figures 1/2/5 chain motivation → architecture → validation, making each module's purpose easy to follow.

Critical weaknesses

1.Limited novelty: the central observation (geometric and peripheral-background invariance) was already proposed and methodized by contemporaneous work that the paper neither cites nor compares against (e.g., GeoProto, arXiv:2605.10885; PSP, MICCAI 2026), so claiming it as the primary contribution does not hold.

2.Unverified critique: the introduction singles out Robust-EMD as failing under drastic appearance change, yet it never appears in the main experiments (Tables 1–2) and is reported only in the appendix.

  1. Mechanism description does not match the implementation: GAI's channel weights are derived solely from the support set and are spatially uniform, contradicting the "dynamic/adaptive" framing, which also undercuts the DGAA comparison.

  2. Ablations do not support the paper's central claim: the largest single gain (+10.40 Dice) comes from an auxiliary BCE loss unrelated to geometry or peripheral background, and the module ablations never state their loss configuration, so per-module contributions cannot be separated.

Questions for authors

1.How does GPS relate to and differ from GeoProto and PSP, which share the same insight and were public before submission, and can you provide direct comparisons?

2.Why are Robust-EMD, MAUP, and Fob-SAM absent from the main tables, and can you add them?

3.What is the loss configuration of each row in Table 4, and what are the net gains of GAI and PPM&PSR once the auxiliary loss is removed?

4.Your retrained baselines differ from the published values by up to 12+ Dice points and your protocol description differs from FAMNet's — how do you explain this?

Reviewer_8ZfkOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3(+1)
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

This paper proposes GPS for cross-domain few-shot medical image segmentation. It encodes support-mask geometry with a dual-ring representation, uses the resulting descriptor to modulate feature channels, and combines foreground and peripheral-background prototypes with prediction refinement. The paper reports improvements on abdominal and cardiac transfer tasks. The appendix adds cross-institution and cross-anatomy evaluations, efficiency measurements, and further ablations.

Critical strengths
  1. The design is coherent for a relevant problem. Combining support-mask geometry with local foreground and background information is a reasonable approach to segmentation under medical imaging domain shifts. The design gives a plausible mechanism for using the limited labeled support information beyond conventional foreground prototype matching.
  2. The empirical evidence of performance gains is substantial. Tables 1–2 report improved mean Dice across all four main transfer directions. The additional cross-institution and cross-anatomy evaluations, together with the component ablations, strengthen the evidence that the benefits extend beyond one transfer pair.
Critical weaknesses
  1. The experimental protocol does not clearly establish the stated task and supervision setting. Section 3.1 requires disjoint source and target classes, but the main experiments list the same anatomical categories across modalities or sequences without explaining how the training classes satisfy this disjointness. It is also unclear whether the training masks are anatomical annotations or generated pseudo-labels. Section 4.2 and the appendix do not fully specify the patient partitions connecting source training, validation, target support, and query evaluation. These details are necessary to judge the claimed novel-class setting, the supervision budget, and whether the evaluation is independent.
  2. The evidence does not establish the claimed invariance properties. Equation (20) aligns directional embeddings of the same mask; it does not directly constrain the representation of a mask to match that of its rotated counterpart. A common but orientation-dependent embedding could satisfy this objective. In addition, Figure 3 does not sufficiently define the distributions and pairing procedure used for the Bhattacharyya coefficients, so it is hard to distinguish within-domain support–query consistency from source–target invariance. This weakens the evidence for the central explanation that geometry and peripheral background provide domain-invariant information.
Questions for authors
  1. Are the training masks anatomical annotations or generated pseudo-labels? Please specify the source and target class sets and the patient partitions for training, validation, support, and query, including whether different cardiac sequences from the same patient always remain in the same partition.
  2. What distributions, normalization procedures, and sample-pairing rules are used for the Bhattacharyya coefficients in Figure 3? Do these measurements compare samples within each domain or directly across source and target domains?
  3. If the rotation-invariance claim is retained, can you measure changes in P_{geo} under rotations of existing held-out masks? If such evidence is unavailable, can the claim be restricted to directional consistency, with an explanation of which conclusions remain supported under that interpretation?
1 comment

Dear Reviewer,

Thank you very much for taking the time to respond to our rebuttal. We hope that our responses and additional experiments have adequately addressed your concerns. If you have any remaining questions or concerns, please let us know. We would be happy to provide further clarifications or additional evidence to address them.

We sincerely appreciate your time and constructive feedback.

Reviewer_7mptOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets cross-domain few-shot medical image segmentation under 1-way 1-shot evaluation, on the premise that organ shape and surrounding-tissue distribution are more stable across domains than appearance. GPS has three modules: GAI encodes the support mask as a dual-ring ray descriptor and uses its similarity to the features to weight channels; PPM mines 10 foreground and 10 peripheral-background local prototypes by farthest-point sampling; PSR fuses global and local similarity maps into the final prediction. Training adds a coarse mask loss L_cor and a rotation-consistency loss L_rotate. Mean Dice is 73.81 / 69.41 on Abd-CT↔Abd-MR and 78.55 / 65.87 on Cardiac LGE↔bSSFP, with lower FLOPs and latency than C-Graph, plus ablations, sweeps, a prostate benchmark and two external datasets.

Critical strengths

The geometry-as-invariant-cue idea is intuitive, and the dual-ring descriptor reasonably represents hollow structures such as LV-MYO.

GPS is markedly cheaper than the strongest baseline: 16.58 GFLOPs and 48.72 ms against 36.66 GFLOPs and 163.52 ms for C-Graph, at 6.52M parameters.

The evaluation spans abdominal and cardiac settings in both directions, a cross-institution benchmark and two external datasets, and the tables are arithmetically consistent.

Critical weaknesses
  1. My main reservation is that the evidence does not yet separate geometry from a generic attention effect. Fig. 3 measures agreement between a support and a query from the same dataset, which shows that shape is stable within a domain — but the same is true of appearance — rather than that it survives the CT–MRI gap, which is the premise of the paper; a source-versus-target measurement, and a control using intensity or texture statistics, would speak to that directly. More importantly, I could not find any experiment in which the geometric descriptor is replaced by a non-geometric one, such as a shuffled mask, a random vector, or a mask-pooled appearance feature. Table 14 comes closest, but DGAA only changes how the geometric vector is applied, so it does not test whether geometry is the ingredient that matters. As the weight is computed from the support mask and then applied to both branches, GAI is functionally support-conditioned channel attention, and its large gain over the base model (59.0 to 70.5 mean Dice) may have little to do with shape. The same ambiguity runs through the ablations: a single ring already clears every baseline on the myocardium by a wide margin (60.4, against a best baseline of 52.9), with the dual ring adding around two points there and under one on the cardiac mean, while a coarse cross-entropy term on the intermediate prediction contributes ten points where the rotation loss contributes one point four. The components that move the numbers are not the ones the paper's story rests on.

  2. Two of the design choices look fragile to me. The rotation loss asks the per-direction embeddings to agree with each other, and that objective is minimised by mapping every direction to the same constant — invariant, but empty of shape; the primary losses may keep this from collapsing in practice, yet nothing in the paper shows the descriptors remain discriminative across organs or directions. In PSR the convolutions run along the axis indexed by sampling order, an index with no anatomical meaning that shifts with the mask being segmented, so the same channel refers to different places in different episodes. I also could not find a clean isolation of PPM: it is only tested on top of GAI, and the refinement network is present in every row of the relevant table, so it is unclear whether the multi-prototype construction or the extra convolutions carry the gain, and a comparison against background prototypes mined from the whole image would settle whether the peripheral definition itself does any work. Finally, the ray construction is not discussed for lobulated masks such as liver or kidney, where keeping only the first and last intersection discards internal boundaries, nor for masks cut off by the image border, nor for the dependence of the whole descriptor on the single support slice.

  3. The gains are uneven, and the protocol makes them hard to place. Most of the improvement comes from the kidneys, while on the liver GPS trails several baselines — five points behind FAMNet in one direction and nine behind RPT in the other — and the CT↔MR average margin is small enough to sit inside the reported standard deviations. All hyper-parameter sweeps are run on a single task, and no validation split is described, so I cannot tell whether the settings were chosen without looking at the target test folds; the sweeps move by several points, which is the same order as the effects claimed for individual components. Encoder pre-training, meanwhile, shifts the result further than any proposed module, which suggests the reported gains lean heavily on the pretrained representation. Single-seed results with no paired tests make these distinctions difficult to settle either way.

  4. Smaller points: the baseline set changes between tables, with RobustEMD and CATNet appearing only in the appendix and QNet and TPM only in the main tables; several cited competitors are never compared; and the abstract, the contributions and the experiments disagree on how many datasets are involved.

Questions for authors

Please refer to the weaknesses.

Comments on the paper

1 comment
Area_Chair_6HM7OpenReview ↗

Please Participate in the Discussion Phase

Dear Reviewers,

The authors have posted their rebuttal and responses addressing your review comments and questions.

Please carefully read their responses (as well as the updated manuscript, if applicable) and actively engage in the reviewer-author discussion phase. If you have any remaining concerns or follow-up questions, please post them so the authors can address them before the discussion period closes.

Your timely participation and updated assessments are essential to helping us reach a fair and well-informed recommendation for this submission.

Best regards,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.