List of Papers Browse by Subject Areas Author List
Abstract
Training-free cross-domain few-shot medical image segmentation aims to segment unseen anatomies without parameter updates, addressing the high cost of dense annotation and domain-specific fine-tuning in clinical practice.
Existing support-driven prompting methods face three limitations: support exemplars are randomly selected without quality assurance, geometric alignment is poorly modeled, and multi-modal prompt capabilities remain underexploited.
We present APEX-SAM, a retrieval-augmented framework with three innovations.
Quality-aware expert retrieval (QAR) builds a dual-stream DINO/SigLIP expert bank with diversity-aware selection to ensure support-query compatibility.
Anatomy-aware prompt mining (APM) performs style-aligned geometric matching and anatomy-guided point sampling from morphological priors.
Hybrid multi-modal fusion (HMF) fuses three SAM branches (point, text, and box prompts) via training-free feature-consensus weighting.
Experiments on three cross-domain benchmarks confirm state-of-the-art performance among training-free methods, with ablations validating each component’s contribution.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0425_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Trump0412/APEX-SAM
Link to the Dataset(s)
CHAOS dataset: https://chaos.grand-challenge.org/
BTCV (Multi-Atlas Labeling Beyond the Cranial Vault) dataset: https://www.synapse.org/Synapse:syn3193805
MS-CMRSeg 2019 dataset: https://zmiclab.github.io/zxh/0/mscmrseg19/
BibTex
@InProceedings{MaoZhi_APEXSAM_MICCAI2026,
author = { Mao, Zhihao AND Chen, Bangpu AND Lei, Qi AND Tan, Jiaqi AND Sun, Kun},
title = { { APEX-SAM: Anatomy-Aware Prompting with Expert Retrieval for Training-Free Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16878},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
1.The problem is clearly defined, with a clear clinical significance. 2.The APEX-SAM framework is not a mere accumulation of individual techniques, but a systematic solution addressing the three major bottlenecks of existing methods: support sample quality, geometric alignment, and multimodal fusion. 3.Make full use of prior knowledge and foundational models
- 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.
Addressing the clinically significant issue of training-free cross-domain few-shot medical image segmentation, a novel retrieval-enhanced framework (APEX-SAM) is proposed. This method integrates three core modules: QAR, APM, and HMF, enhancing segmentation performance in cross-domain scenarios. It achieves significant performance improvements on three public benchmarks, surpassing existing training-free and partially training-required methods. The work demonstrates certain innovativeness and is clearly written. Its main strengths lie in its systematic framework design and notable performance gains; the main concerns are certain dependencies on the concept of “training-free” and the depth of discussion on computational costs. Overall, this is a solid and contributive study.
- 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.
Major Comments: The discussion regarding the premise of being “training-free” should be more in-depth. While the method does not require gradient updates on the target domain, its performance heavily relies on a meticulously curated “expert repository” constructed from annotated source domain data. The process of building this repository (clustering, quality scoring, diversity selection) can itself be viewed as a form of “training” or “meta-training. “ Furthermore, the method for obtaining the crucial parameter q_i is not clearly explained. Additionally, it is necessary to specify the size of the sample repository. The analysis of computational efficiency and scalability is insufficient. What are the retrieval costs and inference speed? The first row (baseline) in the ablation study needs clarification on which prompt was used. It is unreasonable for the second row to appear alone, as your QAR is merely a sample repository and must be used in conjunction with APM and HMF. Furthermore, the very significant performance improvement shown in the final row of results should be analyzed further. Minor Comments: In the abstract, acronyms like QAR and APM should be spelled out in full upon their first occurrence. The authors’ statement regarding SAM 3 contains a misleading characterization: “Third, although modern foundation models like SAM 3 [1] support rich multi-modal prompts, current methods rely on single-modal prompts and lack principled fusion strategies. “ SAM 3 may lack anatomical knowledge. Introducing text prompts into this work could be a source of noise, which the authors need to investigate.
- 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 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.
(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 motivation and writing of this paper are both decent. The reason I gave this score is mainly due to concerns about the authenticity of the ablation experiments, as well as the lack of clarity in many details within the paper. If the author can adequately address these concerns in the rebuttal, I believe the score could be raised to 4.
- 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 author cleared up some of my issues, so I’ve bumped my score up to a four.
Review #2
- Please describe the contribution of the paper
The method proposes a training-free method for binary class segmentation in medical images using a three fold mechansim. It first performs support image selection, followed by multi-modal prompt selection and fusion. The results show impressive improvements on multiple test sets
- 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) Very strong empirical results - Significant gains over existing methods. The experimentation is exhaustive and includes very recent methods
2) Three-way novelty - Addresses poor support selection, geometric alignment and multi-modal prompt fusion.
3) I liked the Chamfer alignment idea for learning a transform of the boundaries
- 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) No analysis on computation time across methods. I would believe that 3 time forward pass of SAM plus other transforms would require much more time than other methods
2) No analysis on the size of the retrieval bank. How many annotated examples would be required for good performance? Say, this method is being used for a new research, a significant initial setup cost would be the creation of a new support bank. Having an ablation over the size of the bank (number of clusters, number of examples per cluster etc) would help.
- 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 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
I would encourage the authors to answer the questions raised in the weaknesses.
- 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 of segmentation in a training-free method. There are methodical novelties and a significant rise in performance. Some questions should be answered, but overall the method is taking the science ahead.
- 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 provide assurance in the rebuttal that the increased results are not caused due to training data leakage, which was a very valid point raised by other reviewers. In addition, I would like the authors have a more comprehensive limitations section in the final paper that acknowledges the increased inference time due to 3 SAM3 forward calls
Review #3
- Please describe the contribution of the paper
This paper proposes APEX-SAM, a training-free, retrieval-augmented framework for cross-domain few-shot medical image segmentation. The method retrieves support examples from an expert bank, aligns them to the query image, generates prompts, and fuses multiple SAM-based prompt branches.
- 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 an important clinical deployment scenario: adapting segmentation to new domains without retraining. The high cost of dense annotation and the difficulty of cross-modality generalization are well motivated. The framework is modular and conceptually appealing. Combining retrieval, geometric alignment, and multi-prompt fusion is a reasonable strategy for training-free segmentation. The use of DINO for structural retrieval and SigLIP for semantic retrieval is plausible. The anatomy-aware prompt generation appears sensible, particularly the use of positive points inside spatially diverse foreground regions and negative points near likely false-positive boundaries.
- 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 largest concern is reproducibility and validity of the foundation-model setup. The method refers to SAM3 and uses text, point, and box prompt branches. However, the cited SAM3 reference is an anonymous under-review work. The paper does not sufficiently clarify what model is actually used, whether it is publicly available, and how text prompting is implemented. This is a major issue for a MICCAI submission because the core method depends on capabilities that may not be accessible or verifiable.
The reported gains are extremely large, especially for training-free cross-domain segmentation. Such results may be possible, but they require unusually careful protocol description. The current paper does not provide enough detail about dataset splits, episode construction, preprocessing, support-label availability, target-label exclusion, and whether the expert bank can contain cases too similar to the query. The “patient-level leave-one-out” description is not sufficient to rule out leakage or overly favorable retrieval.
The comparison with baselines may be unfair or unclear. Several baselines are training-based few-shot segmentation methods, while APEX-SAM uses a retrieval bank, foundation-model prompts, label-filtered retrieval, and possibly an append-only memory update. It is not clear whether all methods receive the same support information, text label, and preprocessed inputs.
The method assumes the query anatomy text label is known. This is reasonable for class-specific segmentation but should be stated more explicitly as a limitation, because it simplifies the task relative to open-set or fully automatic multi-organ segmentation.
The append-only memory update raises concerns. The paper states that inference-time append uses predicted masks and SAM confidence as quality proxies. This could improve reported performance over episodes but may also introduce temporal dependence, confirmation bias, or leakage if not carefully controlled. The ablation shows a large +4.0 pp gain from thresholded append, so the exact evaluation protocol matters greatly. The method is 2D slice-based, while many medical segmentation tasks are volumetric. The paper mentions future 3D extensions, but the clinical relevance would be stronger with 3D consistency, or evaluation on full scans.
- 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 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.
(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 idea is timely and potentially useful, but the current submission has serious concerns around reproducibility, evaluation protocol, and dependence on an under-specified SAM3/text-prompt model. The reported performance is very strong, but the paper needs much clearer experimental controls before the claims can be trusted.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers for recognizing APEX-SAM’s value. We address concerns on protocol, ablation, cost/bank size, fairness, and SAM3/text-prompt reproducibility. Protocol/leakage. The evaluation is patient/volume-disjoint. The original bank is built from source/bank patients never reused as query patients. For each query, all slices from the same patient/3D volume are excluded, preventing adjacent-slice or near-duplicate leakage. Support labels are available only for source entries. Target masks are used only for Dice computation, never for retrieval, bank construction, prompt generation, support selection, or append decisions. In append-only evaluation, the current query is evaluated before any append; appended pseudo-supports contain APEX-SAM predicted masks/confidence scores, not ground-truth masks, and affect only later queries. Bank/fairness. All comparisons use the same benchmark family and source/target splits. APEX-SAM introduces no extra private data; its bank is a compact, quality-filtered subset of the same source pool used by the baselines: 180 curated 2D entries for 7 labels, about 6–7 DINO clusters per label. At inference, each query uses top-L=3 clusters and top-M=5 supports. Retrieval is label-filtered but sample-level: the anatomy name only ensures that a support has a valid mask for the requested structure; after filtering, QAR ranks whole-image supports using DINO/SigLIP, not label-specific optimization. Training-free/QAR details. We agree that “training-free” alone may obscure offline preparation cost. Expert repository construction is data curation/indexing, not meta-training: clustering, quality scoring, and diversity selection operate on a fixed source-support pool and do not learn a transferable model, initialization, or retrieval policy. This setup cost converts random/noisy/redundant supports into compact high-quality memory. The quality score qi is not learned; it is mean frozen-SAM foreground confidence inside the annotated support mask, used only for support ranking/filtering. Ablation. The first row uses the anatomy-name text prompt only, e.g., “liver”. The “QAR-only” row is not a standalone repository: it uses a degraded MAUP-style pipeline where QAR retrieves supports, DINO matching transfers foreground cues, and these cues become point prompts for SAM; APM/HMF are disabled. Thus, it can produce masks and isolates support-bank quality. The final gain comes from optional append-only memory, where high-confidence, non-redundant pseudo-supports improve later retrieval. The fixed-bank QAR+APM+HMF reaches 91.8 Dice, so the core contribution does not depend on append. Cost/scalability. Computation has one-time offline bank construction (DINO/SigLIP keys, clustering, qi scoring, diversity selection) and bounded online inference. Runtime is dominated by frozen SAM3 inference; SAM3 was publicly released by Meta on Nov. 19, 2025, with 848M parameters and a promptable interface, and public reports indicate fast promptable inference, e.g., about 30 ms per frame on H200 for 100+ objects, though not our task-specific benchmark. Additional retrieval, DWT, Chamfer alignment, and fusion are lightweight versus training; APEX-SAM requires no target-domain optimization. SAM3/text/scope. The foundation model is frozen Segment-Anything-3 (SAM3) , with text/point/box branches. Text prompting simply uses the anatomy-name string. We agree SAM3 may lack anatomical knowledge; our claim is a generic multi-modal prompting interface, not SAM3 as an anatomical prior. The text-only baseline is much weaker (72.4), while fixed-bank QAR+APM+HMF reaches 91.8, showing gains from QAR/APM/HMF rather than text alone. APEX-SAM is modular, so SAM3/DINO can be replaced by medically adapted promptable backbones in future extensions. Our claim is limited to class-specific binary 2D segmentation; multi-organ/open-set use and 3D consistency are outside this scope, but they are natural future directions given APEX-SAM’s modular design.
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.
All three reviewers acknowledged that this paper does have some merits. It attempts to address an important problem for practical deployment: adapting segmentation to new domains without retraining. Its experimental gains are strong and outstanding. Its ideas are good and potentially useful. However, the reviewers also raised some major concerns on the manuscript, particularly on the experimental designs.
The reviewers pointed out that the authenticity of the ablation studies should be further elaborated, the computational time and the size of the retrieval bank should be further discussed, the reproducibility of the proposed model should be further demonstrated, the experimental protocol should be more carefully descripted (considering the large gains), the comparisons with baselines should be further clarified, and the dependence on an under-specified SAM3/text-prompt model should be further analyzed.
In rebuttal, please carefully study the comments raised by the reviewers, particularly the above-mentioned major concerns, and provide your responses to these questions.
- 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.
After rebuttal, two reviewers suggest accepting this paper with one reviewer increasing the score from 3 to 4.The other reviewer did not submit the final score after rebuttal. Based on the rebuttal and the comments of the two reviewers, the paper can be accepted.
It is suggested that the authors sufficiently consider the comments raised by the reviewers and further improve the paper in the final version.
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.
The rebuttal adequately addresses the reviewers’ concerns. I agree with R2; the authors should add a more comprehensive limitations section in the camera-ready, acknowledging the increased inference time.
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.
In their rebuttal, the authors have addressed important concerns mentioned by the reviewers regarding setup and data split. Some limitations (e.g., 2-D vs. 3-D setting) and concerns (e.g., preprocessing / runtime / comparability regarding support information) remain to some extent; however, given that two reviewers see rate the paper above acceptance after rebuttal and reviewer #3 has not revised their rating, but their concerns were answered to a substantial extent, I recommend acceptance.