acceptodds
Under review as a conference paper at ICLR 2027

DA-Proto: Distribution-Aligned Prototype Learning for Cross-Domain Few-Shot Medical Image Segmentation

Abstract

Few-Shot Medical Image Segmentation (FSMIS) aims to segment unseen categories with only a few annotated samples. However, domain shifts across imaging devices, protocols, and modalities can severely degrade target-domain performance, motivating Cross-Domain FSMIS (CD-FSMIS). Existing methods mainly exploit cross-domain stable cues in frequency, anatomical structure, or semantic channels, while largely overlooking the support-query foreground distributional relationship. Notably, we observe that support and query images of the same category exhibit consistent foreground distributions within each domain, and this relationship remains stable across domain shifts such as CT MR. Based on this observation, we propose Distribution-Aligned Prototype Learning for Cross-Domain Few-Shot Medical Image Segmentation (DA-Proto), which extracts a stable foreground distribution prior from target-domain support samples and uses it to modulate domain-shifted class prototypes into distribution-aligned prototypes. Extensive experiments across multiple cross-domain settings demonstrate that DA-Proto significantly outperforms state-of-the-art methods.

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

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_qAQKOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses Cross-Domain Few-Shot Medical Image Segmentation (CD-FSMIS), focusing on the prototype representation bias introduced by distribution shifts between the source and target domains. Motivated by the observation that support and query images of the same category exhibit relatively consistent foreground statistical distributions within individual domains, the authors propose DA-Proto, a distribution-aligned prototype learning framework that leverages foreground distribution priors from target-domain support samples to improve prototype adaptation.The method is evaluated on abdominal CT/MRI cross-modality segmentation and cardiac MRI cross-sequence segmentation benchmarks. DA-Proto achieves the highest reported mean Dice scores among the compared methods across all four cross-domain tasks. The authors further provide component and loss ablations, distribution-prior substitution experiments, sensitivity analyses, and computational efficiency comparisons to support the effectiveness of the proposed framework.

Critical strengths

The paper addresses the challenging problem of cross-domain few-shot medical image segmentation. Its attempt to exploit relative distribution relationships between support and query images provides an interesting perspective on domain generalization, although the underlying assumption and methodological novelty require stronger justification.

Critical weaknesses

W1. Insufficient justification of cross-domain distribution stability. Although the authors report high support-query foreground distribution similarity across several domains, these results mainly demonstrate within-domain consistency rather than a domain-invariant relationship. Stronger evidence is needed to establish the claimed cross-domain stability. W2. Limited methodological novelty. The proposed framework primarily combines foreground statistical modeling, reconstruction-based regularization, and additive prototype modulation. The methodological differences from existing prototype adaptation and feature modulation techniques remain insufficiently established. W3. Gap between motivation and method. The motivation emphasizes stable support-query distributional relationships, whereas the proposed method mainly extracts support foreground statistics and adds the learned representation to class prototypes. It is unclear how this operation explicitly achieves distribution alignment or preserves the claimed relationship. W4. Incomplete comparisons with recent methods. Although the paper includes several relevant cross-domain few-shot segmentation baselines, comparisons with more recent methods, particularly those from 2026, would better establish its competitiveness. W5. Limited analysis of the distribution prior. The prior ablations demonstrate the benefit of category-matched information, but it remains unclear whether the improvement specifically comes from distribution modeling or simply injecting additional category-specific support features. Comparisons with simpler support-feature modulation alternatives would strengthen the methodological justification.

Questions for authors

refer to the weakness

2 comments
AuthorsOpenReview ↗

Response to R3

We appreciate the reviewer’s thoughtful feedback and helpful suggestions. Our responses are provided below.

Q1. Clarification of Cross-Domain Stability.

We would like to clarify that the "cross-domain stability" discussed in our paper does not imply that different imaging domains (e.g., CT and MRI) share similar absolute intensity distributions. Instead, our motivation stems from the observation that models trained on a source domain may overfit to domain-specific appearance cues, such as texture and contrast, which can vary substantially across imaging modalities.

In cross-domain few-shot segmentation, both support and query images are sampled from the target domain. The support foreground statistics can provide target-domain-specific information relevant to query segmentation. Our method leverages this information to condition the segmentation prototypes, aiming to reduce reliance on source-domain-specific appearance cues.

Therefore, our method does not assume invariant distributions across domains or identical support–query distributions. Rather, it exploits support-derived target-domain distributional information to improve prototype adaptability under domain shifts. We will clarify this distinction in the revised manuscript.

Q2-Q3. Methodological novelty and the connection between motivation and distribution alignment.

Our key contribution lies in explicitly incorporating support-derived foreground distributional information into prototype learning to improve cross-domain adaptability, rather than relying solely on semantic prototypes or conventional feature modulation.

Specifically, FDM extracts a foreground distribution prototype F_d from the support features and foreground mask. However, directly mapping statistical information into the feature space may not adequately preserve its underlying statistical characteristics. To address this, DFR introduces a statistical reconstruction constraint that encourages F_d to retain the original distributional information. DGPM then integrates the distribution-aware prototype with the semantic prototype to guide query segmentation.

We further clarify that our "distribution alignment" does not explicitly minimize the discrepancy between support and query distributions. Instead, it refers to preserving the statistical information encoded in the learned distribution prototype through reconstruction-based supervision, which subsequently enables distribution-aware prototype conditioning.

Thus, FDM, DFR, and DGPM form a unified framework that connects foreground distribution modeling, statistical information preservation, and prototype conditioning. This design enables the model to leverage target-domain distributional cues for cross-domain few-shot medical image segmentation.

Q4. Comparison with current CD-FSMIS method.

We have additionally compared our proposed DA-Proto with two recent methods, GeoProto and PSP, on the Abdomen CT → Abdomen MR task. The quantitative results are presented in the table below.

MethodLiverLKRKSpleenMean ±\pm Std
GeoProto|72.58 ±\pm 02.51|67.43 ±\pm 04.47|73.87 ±\pm 02.37|58.18 ±\pm 13.90|68.01 ±\pm 04.32
PSP|69.82 ±\pm 06.08|68.52 ±\pm 11.29|78.94 ±\pm 09.78|57.96 ±\pm 18.16|68.81 ±\pm 08.17
DA-Proto|74.56 ±\pm 03.09|73.26 ±\pm 04.30|85.75 ±\pm 04.84|64.98 ±\pm 11.33|74.64 ±\pm 04.41

As shown, DA-Proto consistently outperforms both recent methods across all four anatomical structures (Liver, LK, RK, and Spleen), achieving the best performance in terms individual-organ Dice scores . Specifically, DA-Proto achieves a mean Dice score of 74.64%, outperforming GeoProto and PSP by 6.63 and 5.83 percentage points, respectively.

Q5. Comparisons with simpler support-feature modulation.

We conducted experiments on Abdomen CT →\to Abdomen MR task to compare our DA-Proto with simply injecting additional information (SIAI) of category-specific support features. The results are presented in the table below.

MethodLiverLKRKSpleenMean ±\pm Std
SIAI|69.65 ±\pm 06.47|71.74 ±\pm 05.25|82.41 ±\pm 07.49|63.19 ±\pm 09.72|71.75 ±\pm 05.48
DA-Proto|74.56 ±\pm 03.09|73.26 ±\pm 04.30|85.75 ±\pm 04.84|64.98 ±\pm 11.33|74.64 ±\pm 04.41

As shown, DA-Proto consistently outperforms SIAI across all four anatomical structures, achieving a mean Dice score of 74.64%, which is 2.89 percentage points higher than that of SIAI (71.75%). These results suggest that explicitly modeling and preserving foreground distributional information provides additional benefits beyond simply injecting additional information of support features.

Reviewer_qAQKOpenReview ↗

Reply

Thank you for the additional experiments and clarifications. The rebuttal addresses most of my concerns, particularly regarding W4 & W5. The new results also strengthen the generalization claim.The motivation of the unified framework also lowers my concern on novelty. Overall, I am satisfied with the response and willing to raise my rating.

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

The paper proposes DA-Proto for cross-domain few-shot medical image segmentation. FDM encodes support foreground feature statistics into a distribution representation. A training-only DFR branch regularizes this representation through reconstruction of image-intensity statistics and representation alignment. DGPM then adds the learned representation to foreground prototypes before cosine-based query matching. Experiments cover four transfer directions on abdominal CT/MR and cardiac MRI datasets, with reported improvements in average Dice over the compared methods.

Critical strengths
  1. The reported segmentation results are promising. DA-Proto achieves the highest mean Dice among the compared methods in all four transfer directions, with improvements of 5.68, 7.14, 9.95, and 3.34 percentage points over the respective strongest baselines. These results suggest practical value, subject to clarification of the evaluation protocol.

  2. The method provides a relatively simple way to condition foreground prototypes on support statistics. The component and prior-substitution ablations offer useful evidence that the proposed conditioning contributes to performance, while DFR is used only during training.

Critical weaknesses
  1. The evaluation protocol described in the manuscript does not appear to match the released evaluation code. Section 4.2 and Appendix A.3 describe using one central annotated support slice per target organ to segment all query slices. However, the released scripts set n_part=3, select three annotated support slices from different positions within the organ-containing slice range, and use each support slice for the corresponding query chunk. The query loader also restricts evaluation to slices containing the target organ using ground-truth foreground presence. This resembles an EP1-style evaluation (from ADNet), whereas the manuscript describes single-central-slice support selection. These differences affect the total annotation budget and the use of query foreground-location information. Although the paper states that all methods use the same protocol, the discrepancy makes the actual comparison setting difficult to verify.

  2. The evidence supports useful support-conditioned modulation more directly than the specific explanation of distribution alignment. Figure 1 provides illustrative examples, and Figure 3 provides quantitative evidence of within-domain foreground intensity similarity. However, the sample counts, pairing procedure, and coverage across patients and anatomical positions are insufficiently specified. Moreover, image-intensity similarity does not by itself establish the corresponding property in the learned feature space. DFR is designed to connect these representations, and Tables 3 and 5 provide partial supporting evidence, but the contribution of the proposed distribution information remains difficult to separate from more general support-conditioned feature modulation. This concern also applies to Equation 12, which adds the same vector FdF_d to every foreground prototype. The rationale for interpreting this shared correction as distribution calibration is not sufficiently explained. Its relative magnitude matters for subsequent cosine matching: a sufficiently dominant shared vector could reduce the angular diversity of the normalized prototypes. This is a potential concern rather than an observed failure. A focused control or diagnostic would strengthen the mechanism claim; an exhaustive causal analysis is unnecessary.

Despite the concerns above, the reported improvements across four transfer directions suggest a useful empirical contribution. I consider the remaining concerns potentially addressable through clearer reporting of the evaluation protocol and more focused validation. On balance, I lean toward weak acceptance.

Questions for authors
  1. Evaluation protocol: Which protocol generated Tables 1–2: one central support slice reused for all queries, or three support slices matched to three query chunks as in the released scripts? Please clarify the total number of annotated support slices, whether query foreground-location information is used, and whether every baseline follows exactly the same setting. How should the code be reconciled with Section 4.2 and Appendix A.3?

  2. Figure 3: How many support–query pairs were evaluated, how were the support and query slices selected, and what intensity normalization and distribution-estimation procedure were used? In particular, does the analysis cover different patients and anatomical positions? A comparison against different organs or background regions would help assess how discriminative the observed similarity is.

  3. Prototype modulation: What motivates adding the same FdF_d to all foreground prototypes, and how is its scale relative to the prototypes controlled? A small sensitivity analysis of pk+αFdp_k+\alpha F_d, or measurements of relative norms and prototype angular diversity before and after modulation, would help assess this operation. A capacity-matched support-conditioning control would be another useful way to strengthen the attribution to distribution information.

5 comments
AuthorsOpenReview ↗

Response to Q1 and Q2 of R2.

Q1. Evaluation protocol.

Three support slices matched to three query chunks as in the released scripts generates the results Tables 1–2. Query foreground-location information is used for query slices selection. Every baseline follows exactly the same setting. Details are as follows.

Five-fold evaluation setting are adopted for each dataset. For each fold, we select the third patient (3D volume) as the support set and the remaining patients as the query set. For each target organ, we identify the slices of interest in both support and query volumes using the ground-truth masks. These slices are then divided into three approximately equal-sized partitions (n_part=3 in the code). For each query slice, we select the middle slice of the corresponding support partition as its support image. Thus, each evaluation fold uses three annotated support slices per target organ. Notably, the query ground-truth masks are used only to identify and partition organ-containing slices and are not provided as input to the segmentation model. We have modified the unclear description in the manuscripts.

Q2. Details and addition experiment in Fig 3.

A total of 7,000 support–query pairs were evaluated. Specifically, for each target organ in each dataset (Abd-CT, Abd-MR, Cardiac-LGE, and Cardiac-bSSFP), we first collected all 2D slices containing the target organ. We then randomly sampled 500 support–query pairs, while ensuring that the support and query slices came from different patients. The random seed was set to 2027. The slices were not matched by anatomical position, allowing the analysis to include variations across patients and slice positions. We evaluated four target organs in Abd-CT and Abd-MR and three anatomical structures in Cardiac-LGE and Cardiac-bSSFP, resulting in 7,000 = (4 + 4 + 3 + 3) × 500 pairs.

For intensity normalization and distribution estimation, we first extracted foreground pixels using ground-truth masks and independently for each slice. We then constructed 10-bin foreground intensity histograms and normalized the bin counts by the total number of foreground pixels to obtain probability distributions (P) and (Q) for the support and query slices, respectively. Their similarity was measured using the Bhattacharyya coefficient:

$ BC(P,Q)=\sum_{k=1}^{10}\sqrt{P_kQ_k}. $

In addition, to assess how discriminative the observed similarity is, we added a comparison between support foreground and query background in Fig. 3 (green box plots). For the FG–BG comparison, query background pixels were used instead of query foreground pixels, following the same histogram estimation procedure. As shown, the BC values between support and query foregrounds are consistently higher than those between support foregrounds and query backgrounds across the four domains. This suggests that the observed similarity is not merely due to general intensity distribution overlap, but reflects a certain degree of foreground-specific distribution consistency.

AuthorsOpenReview ↗

Response to Q3 of R2.

Q3.1. Why multiple prototypes use the same F_d.

The multiple foreground prototypes pkp_k capture local semantic features, while the distribution prototype FdF_d represents global category-level information. We therefore add the shared distribution prototype FdF_d to each local prototype pkp_k to incorporate category-level distribution information while retaining local semantic differences.

Q3.2. Sensitivity and prototype diversity of distribution prototype modulation.

We agree that an excessively dominant shared distribution vector may reduce the angular diversity of local foreground prototypes. To investigate this concern, we conduct an inference-time sensitivity analysis by varying λ\lambda in pk+λFdp_k+\lambda F_d on Abdomen CT →\to MR task, while keeping the trained model fixed. We report Mean Dice and the mean pairwise cosine similarity (MPCS) of L2-normalized prototypes, averaged across test episodes.

λ\lambda00.250.50.751.01.251.50
mean Dice ↑|51.67|57.41|65.44|73.16|74.64|66.19| 42.21
MPCS|0.5789|0.5328|0.5083|0.5125|0.5423|0.5883|0.6401

As shown in the table, Mean Dice improves from 51.67% (λ=0\lambda=0) to 74.64% (λ=1\lambda=1), while MPCS decreases from 0.5789 to 0.5423. However, when λ\lambda increases to 1.5, MPCS rises to 0.6401, accompanied by a substantial performance drop to 42.21%. These results suggest that moderate distribution modulation benefits prototype calibration, whereas excessive modulation may reduce angular diversity and impair segmentation performance.

Q3.3. Comparison with generic support-feature injection.

We conduct experiments on the Abdomen CT → Abdomen MRI task to compare DA-Proto with a support-feature injection alternative (SIAI). Specifically, instead of using the DFR-constrained distribution prototype (F_d), SIAI computes the mean and variance of support foreground features and feeds them into an MLP to generate a feature vector for prototype modulation, without the statistical reconstruction constraint imposed by DFR. The same prototype fusion mechanism is retained. The results are reported below.

MethodLiverLKRKSpleenMean ±\pm Std
SIAI|69.65 ±\pm 06.47|71.74 ±\pm 05.25|82.41 ±\pm 07.49|63.19 ±\pm 09.72|71.75 ±\pm 05.48
DA-Proto|74.56 ±\pm 03.09|73.26 ±\pm 04.30|85.75 ±\pm 04.84|64.98 ±\pm 11.33|74.64 ±\pm 04.41

As shown, DA-Proto consistently outperforms SIAI across all four anatomical structures, achieving a mean Dice score of 74.64%, which is 2.89 percentage points higher than that of SIAI (71.75%). These results suggest that explicitly modeling and preserving foreground distributional information provides additional benefits beyond simply injecting additional information of support features.

Reviewer_uuDVOpenReview ↗

For Q1's reply

Please explicitly confirm that the camera-ready manuscript will include the number of support–query pairs evaluated per organ and domain, the slice-selection procedure, coverage of different patients and anatomical positions, and the intensity-normalization and distribution-estimation procedures. These details are essential for assessing the experimental evidence underlying the paper’s motivation. Please also revise the statement, “Notably, we observe that support and query images of the same category exhibit consistent foreground distributions within each domain.”

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

DA-Proto uses foreground statistics from a labeled support image to adjust prototypes for cross-domain few-shot medical segmentation. FDM encodes feature means and standard deviations, while DFR reconstructs intensity statistics and aligns the two representations during training. DGPM adds the learned vector to foreground prototypes for query matching. The model is trained with source-domain superpixel pseudo-labels and evaluated on four abdominal and cardiac transfer tasks.

Critical strengths
  1. The results are the strongest part of the paper. Tables 1–2 report gains of 5.68, 7.14, 9.95, and 3.34 Dice points over the best listed competitor in the four transfer directions. These are substantial improvements across both cross-modality and cross-sequence tasks.
  2. The ablations give useful evidence for DFR: removing it lowers mean Dice from 74.64 to 71.75 (Tables 3–4). This branch is needed only during training, so its gain does not add an inference branch. The appendix also examines support selection and five-shot performance.
Critical weaknesses
  1. The statistics encoder and additive prototype modulation are fairly standard. I see the coupling to intensity-statistic reconstruction as the main addition, but what it offers beyond ordinary statistical conditioning is not yet clear (Sections 3.3–3.5).
  2. Section 3.4 does not explain how the local foreground regions are formed or why the reconstruction uses local rather than whole-foreground statistics.
  3. I could not tell how FDM is used when DGPM is disabled in Table 3. The two no-prior results also need explanation: 55.28 in Table 3 and 51.67 in Table 5.
  4. The liver BC values in Abd-MR vary considerably in Figure 3. It is unclear whether the prior still helps on less similar support–query pairs. Changing the support index, as in the appendix, does not directly answer this.
Questions for authors
  1. What does coupling intensity-statistic reconstruction with the feature prior achieve beyond a standard statistics-conditioned prototype? Which of the existing results support that interpretation?
  2. How are foreground pixels divided into the local regions in Section 3.4? Why reconstruct local statistics rather than statistics over the whole foreground?
  3. How is the FDM output used in the Table 3 rows without DGPM? For the prior controls in Tables 3 and 5, please distinguish retrained models from changes made only at test time, and explain the different no-prior results.
  4. For the lower-BC pairs in Figure 3, does the prior still help, or can it reduce segmentation accuracy? Could you discuss this using existing results or failure cases?
3 comments
AuthorsOpenReview ↗

Response to R1

We appreciate the reviewer’s thoughtful feedback and helpful suggestions. Our responses to the raised concerns are provided below.

Q1. comparison with standard statistics-conditioned prototypes (SSC).

Actually, we have already compared our method with standard statistics-conditioned prototypes (SSC) in Table 3, where the FDM-only configuration corresponds to SSC, improving Mean Dice from 55.28 to 68.08. However, SSC directly maps intensity statistics into the feature space without explicitly preserving distributional information, may generate inadequate discriminative distributional information during the mapping. To address this, DFR introduces a reconstruction constraint to preserve discriminative distributional information, while FDM utilizes it for prototype modulation. Their combination further improves Mean Dice from 68.08 to 70.19.

Q2. Detail formulation for foreground pixels and Why local statistics.

Specifically, we first extract foreground pixels from the support image using the corresponding foreground mask. The extracted pixels are then flattened and sequentially divided into 16 approximately equal-sized groups. For each group, we compute the mean and variance as reconstruction targets for the Distribution Feature Reconstruction (DFR) module.

Compared with whole-foreground statistics, grouped local statistics provide a more fine-grained characterization of foreground intensity distributions, preserving variations that may be obscured by global statistics and encouraging the model to learn richer foreground distribution representations.

Q3. How does FDM work and the difference in Table 3 and Table 5.

FDM without DGPM. When DGPM is disabled, the distribution feature prototype extracted by FDM is directly added to the global foreground prototype. The modulated foreground prototype and global background prototype are then matched with query features via pixel-wise similarity, followed by softmax to obtain foreground probabilities.

In Table 3, each component configuration is independently retrained. In contrast, Table 5 uses the fully trained DA-Proto and only replaces or removes the distribution prior during inference, without retraining. Thus, 55.28 in Table 3 is obtained by an independently trained baseline without our modules, whereas 51.67 in Table 5 is obtained by removing the prior from a trained DA-Proto at inference.

Q4. Discussion of low-BC support-query pairs.

We acknowledge that the BC values for Liver in Abd-MR exhibit greater variability than those of other organs, indicating notable intensity distribution differences in some support–query pairs. Nevertheless, as shown in Table 5, incorporating the distribution prior improves the Liver Dice from 70.12 to 74.56, indicating that the prior is beneficial for Liver overall despite its larger BC variability..

To further address this concern, we have added a low-BC analysis in Appendix C.2. The results show that although segmentation accuracy remains limited in some low-BC cases, DA-Proto still outperforms predictions without the distribution prior in the illustrated examples. As a potential future direction, we will explore distribution-aware support selection strategies to reduce support–query distribution mismatch and improve segmentation robustness.

Reviewer_VKzbOpenReview ↗

Reply

Thank you for the response, which has largely addressed my concerns, particularly the explanation of DFR’s role in Q1 and the clarification of the different ablation settings in Q3. After considering the response and discussion, I maintain my original rating.

We thank the reviewer for the encouraging comments and for acknowledging our clarifications on Q1 and Q3. We appreciate your support and the time you have dedicated to reviewing our paper.

Comments on the paper

1 comment

Dear reviewers,

As the discussion period is limited, please follow up on the latest responses 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.