acceptodds
Under review as a conference paper at ICLR 2027

Elastic Flow-guided Few-shot Medical Image Segmentation with SAM-Refined

10 views

Abstract

Few-shot medical image segmentation focuses on generalizing to unseen categories with limited annotations. However, medical images are characterized by widespread non-rigid deformations and significant intra-class differences, leading to a dilemma of support-query foreground ideal matching. This constitutes a critical factor limiting performance of many existing approaches. To overcome this issue, we propose an Elastic Flow-guided few-shot medical image segmentation with SAM-Refined network (EFG-SAM). First, to alleviate deformation gap between support and query foreground features, we design an Elastic Flow-field Deformation Module (EFDM). Building upon deformed support features, we propose a Deformation-aware Generation Module (DGM) to obtain elastic prototypes containing semantic, details, and background information. Subsequently, leveraging semantic elastic prototypes to suppress background noise in query features, we design an Elastic Localization Module (ELM). Then, we propose an Elastic Aggregation Module (EAM) to achieve efficient segmentation of query features under the guidance of these elastic prototypes. Finally, during the inference stage, SAM is seamlessly integrated into EFG as a training-free enhancement component to improve final segmentation results. Extensive experiments on three datasets show that our method outperforms state-of-the-art approaches by 2.45% on average and by 8.30% in cross-domain settings, demonstrating its effectiveness in bridging the deformation gap.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Elastic Flow-field, Few-shot Medical Segmentation, Deformation-aware Generation Module

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_1unMOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper adress FSMIS by expanding a single support feature through random elastic deformation, generating multiple features with different morphologies for prototype extraction. In this way, it provides more candidate representations for support-query matching and attempts to mitigate the impact of support-query bias. Despite the seemingly strong experimental results, I find that the proposed design is not sufficiently justified and, in my understanding, contains several questionable aspects, including but not limited to random deformation sampling, the computational overhead, and the lack of clear validation of the claimed advantages. Also, I find that the introduction of SAM in this work seems more like an add-on to make performance look better, without any dedicated design. Currently, I am inclined to reject this paper.

Critical strengths
  1. The support-query bias (or intra-class differences) addressed in this paper is indeed a real issue in FSMIS.
  2. The proposed method achieves strong experimental results.
Critical weaknesses
  1. Support-query bias appears to be an old research problem in FSMIS and has already been considered even in very outdated works such as Q-Net and RPT. I do not see sufficiently valuable new insights from this work.

  2. The overall method design is not elegant and, in my view, tends to complicate a relatively simple problem for limited improvement. For example, replicating the support features multiple times and processing all of these copies can introduce substantial computational overhead. Even the authors themselves acknowledge in the limitation section that the increase in computational cost and memory consumption is significant. Therefore, I am not convinced that this method would have a practical advantage in real-world deployment.

  3. In EFDM, the deformation is randomly sampled from a uniform distribution. My understanding, based on the explanation in Appendix A.1, is that the authors assume that there is a certain probability that the query feature can be approximated by applying a deformation to the support feature. If this is the case, why does increasing the number of sampled deformations actually decrease the final performance, as shown in Figure 6? Intuitively, sampling more deformations should increase the probability that at least one deformed support representation is close to the query.

  4. In my view, poor support-query matching is more likely to be caused by semantic feature misalignment than by spatial morphological differences. I am therefore confused about why a random and non-learnable spatial deformation can effectively address this issue. I suggest that the authors conduct additional experiments to separately investigate the effects of semantic discrepancy and spatial/geometric discrepancy on support-query matching.

  5. The construction of the local prototypes is somewhat different from previous methods such as RPT and GMRD. The authors claim that averaging along rows and columns captures vertical texture variations and horizontal structural consistency. I would be interested in seeing experiments that directly validate this claim, particularly comparisons showing why this row/column-based construction is better than existing local partitioning or local prototype construction strategies.

  6. How are the prompts used for SAM extracted from the coarse prediction? In addition, I think FoutF_{out} in Eq 6 may be better denoted by MM, since FF has already been extensively used throughout the paper to represent feature maps.

  7. Section 3.5 explains the effectiveness of the proposed method by saying that it reduces the L2 distance between support and query features. However, the actual matching operations throughout the method are based on cosine similarity. The authors should therefore additionally report how cosine similarity changes before and after deformation, since cosine similarity, rather than L2 distance, is the metric that actually determines the matching results.

  8. Since this work still follows a coarse prediction -> SAM -> refined prediction pipeline, I am interested in the performance before SAM refinement and in the results of using other existing FSMIS models as prompt generators for SAM, since any segmentation model could potentially provide prompts for SAM. What is the specific advantage of EFG-SAM and what design is truly necessary for supporting SAM?

Questions for authors

Please see weakness section.

Reviewer_oMB1OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents the EFG-SAM framework for few-shot medical image segmentation (FSMIS The authors specifically design the framework to alleviate the effects of large intra-class variations caused by non-rigid elastic transformations, because soft tissues in the experimental datasets don’t maintain a fixed shape in practice. At inference, SAM is added as a training-free refinement step. At inference, SAM is added as a training-free refinement step. The proposed EFG-SAM has been evaluated on relevant benchmark datasets, showing performance gains over the baseline methods.

Critical strengths
  1. A clear practical motivation. Support-query deformation and intra-class variation are real issues in FSMIS.
  2. Reported results against the baseline methods look good.
Critical weaknesses
  1. The technical contribution is incremental, and it is unclear where the improvement comes from. The core module EFDM is basically elastic augmentation applied in feature space. Looking at Table 12, SEA already reaches 84.88 against 85.68 for the proposed EFDM, only a marginal 0.8 Dice gain.
  2. SAM is used to refine every prediction at inference, but no ablation shows the results without SAM. It is therefore unclear how much of the improvement comes from the proposed modules rather than from SAM.
  3. There are many inconsistencies between the text and the tables. For example, Section 4.2 describes the cross-domain experiment as Abd-CT → Card-MRI, but Table 3 reports CT → MRI and MRI → CT. Also, in Section 4.3, the text says the baseline is 80.29% and rises to 84.13% with EPG (EFDM and DGM), while Table 4 shows 79.55 and 84.19. In addition, the offset value 0.05 is said to be tested, but Table 5 only shows learnable, 1, 0.5 and 0.1.
Questions for authors
  1. Can the authors report the results of EFG-SAM without SAM refinement, on all three datasets and both settings? Please also compare with other methods under the same condition (all with SAM or all without SAM).
  2. Where are the results for methods such as GFSAM, FSSAM and AdaptiveFSS etc. mentioned in Section 4.2?
  3. The manuscript needs careful proofreading; many sentences are grammatically incorrect, and citations should be in an author–year style rather than numbered, as required by the ICLR template.
Reviewer_UgC9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EFG-SAM for one-way one-shot medical image segmentation, motivated by non-rigid deformation and intra-class variation between support and query anatomy. EFDM jointly deforms support features and masks without learnable deformation parameters. DGM extracts global, local, boundary, and background prototypes, which ELM and EAM use for localization and aggregation. SAM refines the prediction during inference without additional training. Experiments cover abdominal MRI, abdominal CT, and cardiac MRI, with component and hyperparameter ablations, cross-domain evaluations, and supplementary analyses of shape complexity, semantic fidelity, and feature distances.

Critical strengths
  1. The motivation is relevant and the design is concrete. Jointly deforming support features and masks provides a practical way to enrich the representations available from a single annotated example. Incorporating local, global, boundary, and background information is a useful response to anatomical variation, and EFDM adds no learnable deformation parameters.

  2. The reported empirical results are promising. Tables 1-2 show the highest mean Dice among the reported baselines in all five within-domain configurations. The prototype ablations and supplementary comparisons with other deformation methods provide additional evidence of practical value. These results support my weak-acceptance recommendation, while the specific source of the gains deserves clearer characterization.

Critical weaknesses
  1. The benefit specifically attributable to deformation-induced boundary consistency is not sufficiently isolated. Fig. 8 measures support-query mask overlap, Table 10 describes organ shape complexity, and Table 13 measures feature L2 distances. These analyses are useful, but they do not directly quantify query boundary accuracy. Table 7 shows Dice gains from boundary prototypes, yet this does not separate the effects of deformation, prototype construction, the boundary loss in Appendix C.8, and SAM refinement. A matched comparison using boundary-sensitive evaluation before and after SAM would make the central contribution more convincing.

  2. Information preservation and inference cost need clearer characterization. Table 11 reports improved mean support Dice (86.24 to 88.39), but spleen performance drops from 89.09 to 82.94, indicating organ-dependent effects. Resampling and L2-norm-based Top-K selection may change which fine or low-response details are retained; systematic information loss or selection bias has not been established either way. Table 2 already reports 5.86 FPS, compared with 19.03 for DSM and 4.25 for FoB+SAM. However, it does not separate deformation, prototype generation/selection, decoding, and SAM costs. The current evidence therefore cannot identify whether prototype filtering is a substantial runtime bottleneck.

Questions for authors
  1. Can you quantify EFDM's contribution to query boundary accuracy using Boundary F1, HD95, or ASD, with the prototype budget, decoder, boundary-loss setting, and SAM configuration held fixed? Results before and after SAM, and a comparison with and without boundary prototypes, would help distinguish boundary improvements from overall Dice gains.

  2. How is semantic fidelity measured in Table 11, and what explains the spleen drop from 89.09 to 82.94? Please clarify whether the original undeformed support representation is retained. Can identity-only versus identity-plus-deformed features, together with a controlled comparison of prototype selection rules, show whether fine boundary details or particular anatomical regions are systematically discarded?

  3. Can you report inference time for EFDM, prototype generation and Top-K selection, decoding, and SAM separately, along with peak inference memory under the same hardware and input settings? Please clarify whether Table 2 includes SAM and show how runtime changes with deformation count and prototype budget, including the measured benefit of Top-K selection relative to iterative clustering.

Comments on the paper

1 comment

Dear reviewers and authors,

As the discussion period is limited, please follow up on the latest responses and comments in a timely manner to address any remaining concerns. 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.