List of Papers Browse by Subject Areas Author List
Abstract
Few-Shot Medical Image Segmentation (FSMIS) offers a powerful solution to data scarcity but struggles to generalize across different imaging modalities (e.g., CT to MRI). This performance collapse stems primarily from the drastic texture discrepancies between domains, which mislead models trained on source-specific intensity distributions. While existing methods attempt to align frequency or local texture features, they often fail to decouple semantic structure from domain-specific appearance. To address this, we identify a critical invariance: despite distinct imaging physics, the position and geometric shape of organs remain robustly consistent across modalities. Therefore, we propose PSP, a novel framework that harnesses Position and Shape Priors for cross-domain medical image segmentation. Specifically, PSP first introduces a Position Coordinate Embedding (PCE) module to inject relative spatial coordinates for rapid organ localization. Subsequently, a Shape Prototype Modulation (SPM) module constructs domain-invariant structural prototypes via explicit shape priors, effectively filtering out texture noise. Furthermore, the Hybrid-Prototype Prediction (HPP) module adaptively calibrates the support prototype to the query feature distribution, mitigating feature misalignment. Extensive experiments on two public medical imaging datasets demonstrate that PSP significantly outperforms state-of-the-art methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1456_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/xubin471/PSP
Link to the Dataset(s)
N/A
BibTex
@InProceedings{XuBin_PSP_MICCAI2026,
author = { Xu, Bin AND Zhu, Yazhou AND Zhang, Haofeng},
title = { { PSP: Harnessing Position and Shape Priors for Cross-Domain Few-Shot Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes PSP, a cross-domain few-shot medical image segmentation framework that leverages spatial location and shape priors of target objects. Specifically, the proposed PCE module integrates feature representations with their spatial positional information, enhancing the model’s awareness of anatomical structures. The SPM module exploits shape priors derived from FFT and SDM to modulate support prototypes. Furthermore, the HPP module refines the support prototypes in a coarse-to-fine manner. The collaboration of these modules effectively improves cross-domain segmentation performance.
- 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 addresses the cross-domain few-shot segmentation problem by leveraging anatomical structure stability and enforcing shape consistency between support and query images. (2) Ablation studies demonstrate the effectiveness of the proposed modules, and the overall method achieves competitive performance.
- 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) Few-shot segmentation is typically designed to generalize to unseen classes. However, the proposed positional prior appears to impose certain constraints on both training and testing categories. From the experimental setup, both training and testing are conducted on different modalities of the same dataset. It is unclear whether the learned positional and shape priors would remain effective when the test classes differ significantly from the training ones. For example, if the model is trained on abdominal organs but evaluated on chest X-ray or endoscopic images, would such priors still be valid? （2）In the SPM module, instead of directly fusing P_prior with P_s to obtain P_s’, the authors use the similarity matrix 𝑤 between P_prior and F_seq to modulate P_s. What are the advantages of this design choice compared to a more direct fusion strategy? (3) It would be beneficial to evaluate the proposed method on target domains that differ more significantly from the training data (e. g. , Chest X-ray) to further validate the effectiveness and generalizability of the positional and shape priors in few-shot segmentation.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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.
(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?
The effectiveness of the proposed positional and shape prior assumptions in domain generalization settings is acknowledged. However, few-shot segmentation inherently requires generalization to unseen classes during testing. The validity of these assumptions under such a setting is therefore questionable, and this concern constitutes a key factor influencing my overall rating.
- Reviewer confidence
Very confident (4)
- [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.
N/A
Review #2
- Please describe the contribution of the paper
This paper proposes PSP (Position and Shape Priors), a framework for cross-domain few-shot medical image segmentation (CD-FSMIS). The key motivation is that while imaging textures vary significantly across domains (e.g., MRI vs. CT), anatomical position and shape remain relatively invariant. To leverage this, the authors introduce three modules: Position Coordinate Embedding (PCE), Shape Prototype Modulation (SPM) and Hybrid-Prototype Prediction (HPP). Experiments on cross-modality (CT↔MRI) and cross-sequence cardiac MRI datasets show consistent improvements over prior methods, achieving new state-of-the-art performance.
- 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 paper addresses a well-motivated problem by identifying a key limitation of existing CD-FSMIS methods, namely their over-reliance on texture features. The main insight, that position and shape can serve as domain-invariant priors, is intuitive, clinically meaningful, and clearly justified. The framework design is innovative and effective, with a modular structure following the sequence PCE -> SPM -> HPP. In particular, the Shape Prototype Modulation (SPM) module is well-designed, incorporating explicit shape descriptors into prototype learning. Overall, the framework aligns conceptually with the stated hypothesis. Empirically, the method demonstrates strong performance, achieving state-of-the-art results in cross-modality segmentation between CT and MRI, as well as cross-sequence cardiac MRI. The gains are consistent and meaningful, with Dice score improvements around 3–7%. The ablation studies are solid, providing clear validation for each component. The shape priors implemented in SPM contribute the largest improvement, and the analysis of hyperparameters, such as the number of Fourier components, is informative and useful.
- 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.
The methodological novelty of the paper is limited. Most of its components are based on existing techniques: coordinate embeddings are widely used, shape descriptors such as SDM and Fourier are well-established, and prototype adaptation is common in few-shot segmentatio. The main contribution lies more in the integration of these elements rather than in introducing fundamentally new methods. The paper focuses on a single support image formulation and does not clearly describe how the method extends to the more general K-shot setting. In particular, it is unclear how shape priors and prototype modulation are aggregated across multiple support examples, which is non-trivial given the explicit shape encoding design. The proposed method is formulated in 2D and does not explicitly address how it extends to 3D volumetric data, despite experiments being conducted on 3D datasets (CT/MRI). In particular, the position encoding (polar coordinates) and shape modeling (2D Fourier descriptors) are inherently 2D, and it is unclear how these components generalize to volumetric settings. Clarification on whether the method operates slice-wise or incorporates 3D spatial context would strengthen the paper.
- 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.
(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 paper presents a well-motivated and practically impactful extension to cross-domain few-shot segmentation by incorporating explicit position and shape priors. While the individual components rely on established techniques, their integration into a unified prototype-learning framework is novel in this context and leads to consistent performance gains. The contribution is therefore incremental but meaningful, rather than fundamentally groundbreaking.
- 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.
The authors have adequately addressed the main concerns raised during review. In particular, the rebuttal clarifies that the contribution is not merely a combination of existing modules, but a principled framework leveraging anatomical consistency (position and shape) as robust cross-domain priors for FSMIS. The additional 5-shot experiments demonstrate that the proposed shape-aware prototype aggregation generalizes beyond the 1-shot setting and yields meaningful performance improvements. Furthermore, the clarification regarding the slice-wise 2D implementation is reasonable. Overall, despite moderate methodological novelty, the paper presents a well-motivated, technically solid, and empirically strong contribution with clear relevance to MICCAI.
Review #3
- Please describe the contribution of the paper
1.A novel framework for cross-domain FSMS that explicitly leverages organ shape and position as domain-invariant priors. 2.Use of spectral shape descriptors derived from 2D boundary contours via DFT, providing a multi-scale representation that is theoretically invariant to translation, scale, and in-plane rotation. 3.A prototype prediction strategy used to align support set prototypes with the query distribution.
- 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 motivation to use shape and position as domain-invariant priors for FSMS is well-grounded, as these features are generally more robust than intensity distributions across modalities (CT/MRI). 2.The use of spectral descriptors (DFT) is a technically sound way to achieve translation, scale, and rotation invariance for boundary contours. 3.The method demonstrates consistent Dice score improvements over several baseline organs in cross-modality and cross sequence evaluation tasks.
- 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.Ambiguity in Latent Dimensions: The variable Cs is introduced in Equation (4) without a clear definition of its relationship to the primary latent dimension C, why different dimensions? 2.Undefined Few-Shot Protocol: The paper fails to state the “shot” count (k), the size of the support set and the selection process of the support set. 3.It is unclear if baselines were retrained or if results were cited from prior work. Furthermore, the selection process for the support set isn’t stated, raising concerns about whether the support sets were consistent across all compared methods. 4.Evaluation relies on Dice scores alone. Given focus on shapes and boundaries, metrics such as HD95 and ASD should be included. 5.No statistical significance testing was provided. 6.The authors don’t discuss failure cases or limitations. For example, Figure 3 shows empty regions within the liver segmentation while some baselines don’t. 7.Although an anonymized link to the source code is provided, some implementation details are missing, like the size of the dimensions C and Cs, Cross/Self attention number of heads and number of layers.
- 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
1.What is the definition of Cs, and why is it different from C? 2.Regarding the cross-attention in the prototype fusion: is this multi-head attention? Please provide the number of heads and layers used. 3.How was the support set selected for each task, and did competing methods use the exact same slices? 4.How many slices were used for training versus testing in the CT and MRI datasets? 5.Why does the model appear to produce empty regions in large organs like the liver, as seen in Fig. 3? 6.What was the specific k-shot setting and support set size used for the reported results? 7.In the SPM module - what is the contribution of geometric statistics features vs. the spectral shape descriptors?
- 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?
The paper presents an interesting approach to cross domain few shot medical image segmentation by focusing on shape and organ position. However, the paper contains significant documentation gaps, most notably the missing k-shot count and support set details, and overall evaluation protocol details making it hard to determine if the SOTA claims are fairly benchmarked.
- Reviewer confidence
Very confident (4)
- [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.
The authors addressed most of my major concerns, including the latent dimensionality, attention configuration, k-shot setup, support selection protocol, additional boundary metrics, and statistical significance testing. Some issues remain, like the lack of a clear analysis separating geometric statistics from spectral shape descriptors in SPM, and the exact slice counts should be stated more explicitly.
Author Feedback
(R2) The anonymous code link is explicitly included in the last sentence of the Abstract. (R2, MR) We clarify that while the absolute position and shape vary across organs, the relative position and shape consistency between support and query images remains the core mechanism of our PSP. As a prototype-based method, PSP dynamically extracts positional and shape information from the support image and mask at inference time. It does not memorize fixed coordinates and shape but instead learns a generalized mechanism to align priors, enabling robust generalization to unseen classes. We evaluated PSP (trained solely on Abdominal CT and inference on Chest X-ray), PSP achieved an average Dice of 75.71%, significantly outperforming FAMNet (69.80%) by 5.91%. These results demonstrate our powerful generalization to unseen classes, independent of training categories and imaging protocols. (R2, MR) SPM utilizes the similarity matrix w to dynamically modulate P_s. In contrast to direct fusion with static learned weights, w enables the model to automatically identify shape-relevant channels during inference, maintaining its effectiveness even under significant domain shifts. Furthermore, this similarity-based modulation is parameter-free, making it more robust against overfitting than adding convolutional layers in data-scarce few-shot scenarios. (R3) Our novelty is not a mere “stacking” of components, but a innovative leverage of the unique anatomical consistency (position and shape) of medical images. As the Meta-Reviewer recognized, this approach is “intuitive and effective”. The resulting SOTA performance on challenging cross-domain tasks proves that our integration provides a non-trivial contribution far exceeding a simple accumulation of existing techniques. (R3, MR) PSP extends to 5-shot by averaging modulated features across 5 independent support samples. Since all (5-shot) support samples represent the same organ, this aggregation builds an anatomical shape consensus while suppressing slice-specific noise. Specifically, on the Abd-CT to Abd-MR task, our 5-shot performance achieves a mean Dice of 72.49% (Liver 75.29, RK 81.84, LK 71.80, Spleen 61.04), representing a 3.13% gain over the 1-shot setting (69.36%). This 1-shot to 5-shot gain confirms that PSP robustly benefits from K-shot settings by effectively aggregating dynamic anatomical shape priors from support samples. (R3, MR) PSP operates slice-wise per FSMIS standards for fair comparison and to leverage 2D pre-trained backbones. It is inherently extensible to 3D (e.g., via xyz coordinates, aggregate shape features across multiple support slices), though fully 3D modeling remains a future goal limited by current data scarcity. (R4, MR) C means Channel. $C$=512 for Feature, $C_s$=256 for Shape. No relationship of them. (R4, MR) For fairness, all experiments strictly follow FAMNet’s 1-way-1-shot protocol with 5-fold cross-validation: for each epoch, the third 3D volume serves as the support set and the rest as queries. (R4, MR) Cited results for identical protocols, retrained other baselines. (R4) Our PSP (mean Dice 69.39%, HD95 31.05mm, ASD 10.27mm) outperforms FAMNet (Dice 65.79%, HD95 56.78mm, ASD 18.18mm). (R4) A paired Wilcoxon signed-rank test on per-case/per-organ Dice (p=0.0087< 0.05) confirms PSP’s improvement over FAMNet. (R4, MR) As shown in Fig.3, baseline methods (FAMNet, DSM) avoid holes through over-smoothing, which results in severe under-segmentation (missing over half the liver). In contrast, PSP preserves the complete geometric structure despite localized holes. These holes are easily rectified by standard post-processing, whereas the structural information lost by baselines is irrecoverable. (R4, MR) Both Cross/Self-attention modules are configured with 1 head and 1 layer. This minimalist design is specifically intended to mitigate the risk of overfitting and ensure robust feature alignment under the extreme data scarcity typical of CD-FSMIS scenarios.
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 genuinely appreciated your paper’s well-grounded clinical motivation and found the integration of position and shape priors to be an intuitive and effective approach for cross-domain segmentation, leading to consistent performance improvements. To build a strong rebuttal, focus on clarifying the specific documentation and methodological gaps they identified, without the need to run new experiments. Please explicitly detail your evaluation protocol—such as the k-shot settings, support set selection process, and baseline consistency —and clarify implementation specifics like the latent dimensions ($C$ vs. $C_s$) and how your 2D formulation extends to K-shot and 3D volumetric data. Additionally, providing a clearer rationale for your SPM fusion design choices , discussing the method’s generalization capabilities when evaluating on unseen classes , and explaining observed visual anomalies (like the empty regions in the liver segmentation) will gracefully address their core concerns and strengthen your submission.
- 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.
The authors have provided clear and convincing answers to the technical questions raised in the rebuttal, including $k$-shot settings, baseline consistency, and the extension to 3D volumetric data. The authors have successfully justified their core design choices and demonstrated the methodical rigor of their framework, proving its robust generalization across unseen classes. Consequentially, the reviewer has raised the rating from “Weak Reject” to “Accept”. Therefore, the AC recommends the acceptance of this paper.
Meta-review #2
- 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 is a borderline paper with mixed reviews. Reviewers agreed that the paper addresses a well-motivated problem by leveraging anatomical position and shape as domain-invariant priors for cross-domain few-shot medical image segmentation, and that the proposed SPM module and overall framework achieve consistent and meaningful performance gains. However, they also raised serious concerns about the paper’s soundness and significance: (1)the methodological novelty is moderate, as most components are built on established techniques, with the main contribution lying in their integration rather than the introduction of fundamentally new concepts; (2) the evaluation is limited in scope, focusing primarily on 2D slice-wise processing without a clear pathway to 3D volumetric generalization, and lacking validation on domains substantially different from the training data; and (3) several experimental details—such as latent dimensions, few-shot protocol, support set selection, and statistical analysis—were initially unclear or incomplete. The rebuttal clarified key aspects, but some residual ambiguities remain. After careful deliberation, I recommend accepting this manuscript, as the core idea of explicitly incorporating shape-aware geometric priors into few-shot segmentation is both conceptually sound and empirically effective, offering a valuable direction for improving cross-domain robustness in medical imaging.