List of Papers Browse by Subject Areas Author List
Abstract
Source-Free Domain Adaptation (SFDA) is imperative for privacy-preserving medical image analysis, allowing models to adapt to target domains without accessing source data. However, existing SFDA methods face fundamental limitations: entropy-based approaches (e.g., TENT, SAR) only update limited model parameters and suffer from instability; self-training or contrastive consistency methods (e.g., CCRC) still rely on the source model for pseudo-label generation, inevitably accumulating noise under domain shift; and even recent SAM-based correction (e.g., IPLC) depends on a single foundation model without semantic-level verification. In this paper, we propose RaMA, a Reliability-Aware Multi-Agent Adaptation Framework that introduces Large Language Models (LLMs) into the SFDA loop. Specifically, we leverage a Multimodal LLM as a Cognitive Meta-Controller to perform semantic error checking and generate spatial constraints, which guide a Multi-Agent System (MAS) of diverse SAMs to produce consensus-based pseudo-labels. Furthermore, an LLM-derived reliability score dynamically weights the self-training loss, suppressing noisy corrections from propagating. Extensive experiments on the M\&Ms cardiac segmentation benchmark show that RaMA outperforms all compared SFDA methods by a large margin, improving average Dice by 5.58% and reducing ASSD by 50% over the best existing approach.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4948_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/TianyuZhang-hdu/RaMA
Link to the Dataset(s)
M&MS dataset: https://www.ub.edu/mnms/
BibTex
@InProceedings{ZhaTia_SourceFree_MICCAI2026,
author = { Zhang, Tianyu AND Ji, Jianhang AND Cheng, Zhiming AND Zhao, Jianxiang AND Xiao, Fengjun AND Wang, Shuai},
title = { { Source-Free Domain Adaptation for Medical Image Segmentation via LLM-Agent Collaboration } },
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
1.This paper proposes RaMA, a new SFDA framework that integrates the Multimodal LLM into medical image segmentation. 2.RaMA utilizes an LLM to perform semantic error checking and translate anatomical reasoning into point prompts for the SAM series to generate masks. 3.To mitigate noise propagation, RaMA employs dynamic loss weighting based on the rectified confidence derived from both the LLM’s reasoning and SAMs’ outputs.
- 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.This is a pioneering attempt to leverage the perception capability of MLLM to guide the segmentation task. 2.Utilizing the reliability metrics from both the MLLM and SAM series to weight the training loss, to avoid divergence caused by suboptimal source model predictions.
- 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.It is recommended to provide a figure that contrasts the source-model prediction with the pseudo-labels revised by the proposed MLLM+SAM module to explicitly illustrate the quality enhancement. 2.The training phase entails substantial computational overhead, because each forward pass needs one MLLM inference and three SAM inferences.
- 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.
(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?
This paper provides a novel approach to combining MLLMs and medical image segmentation, leveraging the perceptual capabilities of MLLMs. Also, the experiments show that this framework can achieve superior performance compared to other methods (e.g., about a 10% DSC improvement on Center C).
- 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.
After reading the authors’ rebuttal, I believe they have adequately addressed my main concerns. Overall, I remain positive about the work and support acceptance.
Review #2
- Please describe the contribution of the paper
The paper proposes RaMA, an SFDA framework for medical image segmentation that introduces an LLM-guided multi-agent refinement pipeline into pseudo-label adaptation. Specifically, it uses a multimodal LLM to identify semantic segmentation errors and generate spatial prompts, leverages multiple SAM-based agents to produce consensus-refined pseudo-labels, and applies reliability-aware weighting during self-training to reduce the impact of noisy corrections.
- 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 explores an interesting and timely direction by introducing LLM-guided semantic checking into SFDA for medical image segmentation. Using a multimodal LLM to identify semantic errors and convert them into spatial prompts is a relatively novel and creative design. 2.The overall framework is reasonably well structured. The combination of LLM-based spatial constraint generation, multi-agent SAM refinement, and reliability-aware self-training forms a coherent pipeline for improving pseudo-label quality under domain shift.
- 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.The paper lacks qualitative visualization of segmentation results. Since the method emphasizes semantic error correction and pseudo-label refinement, visual comparisons would be important to verify whether the refined masks are anatomically more plausible and have better boundaries. 2.The paper does not report standard deviations or statistical significance analysis when comparing with competing methods, which makes the robustness of the reported improvements less clear. 3.Given the important role of SAM3, MedSAM2, and SAM-Med2D in the proposed framework, stronger comparison experiments with these models on the same datasets would make the evaluation more convincing and help clarify how much improvement comes from the RaMA framework itself. 4.The ablation study is still not sufficiently comprehensive. In particular, it does not clearly isolate the contribution of the LLM component, the individual SAM agents, or the separate terms in the reliability weighting design. 5.The necessity of the LLM component is not fully justified. The paper does not compare against simpler alternatives for generating corrective prompts, such as heuristic or uncertainty-based strategies, so it remains unclear whether the gain truly comes from semantic reasoning by the LLM.
- 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?
I am inclined to give a weak reject at this stage, mainly because the current experimental evidence is not yet sufficiently comprehensive and the contribution of several key components is still not fully disentangled. My final recommendation will depend on whether the authors can clarify these issues in the rebuttal.
- 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.
After reading the rebuttal, I am willing to support acceptance. The authors clarified the direct inference baseline, reliability design, and the role of the MLLM. They also promised to add visualization, statistical analysis, and reproducibility details, which partially addresses my concerns. Some limitations remain, but the main issues are reasonably clarified.
Review #3
- Please describe the contribution of the paper
The paper proposes RaMA (Reliability-Aware Multi-Agent Adaptation), a Source-Free Domain Adaptation (SFDA) framework for medical image segmentation that integrates Large Language Models into the adaptation loop. The main contributions are:（1）LLM-driven spatial constraint generation: A Multimodal LLM (Qwen3-VL) is used as a “Cognitive Meta-Controller,” prompted to act as an expert radiologist that inspects initial pseudo-labels and outputs structured spatial constraints — positive/negative point coordinates indicating False Negative and False Positive regions — translating semantic-level reasoning into geometric prompts compatible with promptable segmentation models. （2）Heterogeneous Multi-Agent System: Three diverse SAM variants (SAM3, MedSAM2, SAM-Med2D) independently produce candidate refined masks guided by the LLM-generated constraints, which are then fused via pixel-wise majority voting to mitigate single-model hallucinations and leverage complementary inductive biases (general boundaries, medical priors, slice-wise consistency). （3）Reliability-aware self-training: A sample-wise reliability score combining LLM semantic confidence and inter-agent structural consistency is used to weight the self-training loss, suppressing noisy refinements from propagating through iterative adaptation rounds.
- 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.Novel and timely framing. To the best of my knowledge, this is one of the first works to integrate a Multimodal LLM as a semantic-level controller into the SFDA loop for medical image segmentation. Existing SFDA methods rely on entropy minimization or self-training with pseudo-labels generated by the source model alone, which is fundamentally limited under strong domain shift. Using an LLM to “diagnose” segmentation failures and translate semantic reasoning into geometric prompts (FP/FN points) is a creative formulation that opens an interesting direction for the community. 2.Heterogeneous multi-agent design. Combining SAM3 (general boundaries), MedSAM2 (medical priors), and SAM-Med2D (slice-wise consistency) and fusing their outputs via majority voting is a sensible way to mitigate single-model hallucinations and exploit complementary inductive biases. This is a genuine improvement over IPLC’s reliance on a single SAM variant. 3.Reliability-aware weighting concept. The idea of combining LLM-derived semantic confidence with inter-agent structural variance into a sample-wise reliability score for soft-discarding unreliable pseudo-labels is well motivated, and addresses a real problem (confirmation bias and error accumulation) in self-training-based SFDA.
- 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.A key zero-shot baseline is missing. The paper does not report what happens if “LLM + Multi-SAM consensus” is applied directly at inference time without any target-model self-training. Without this, it is unclear whether the SFDA self-training loop actually contributes, or whether inference-time prompting alone explains most of the gains. This is essential to validate the paper’s framing as an adaptation method. 2.Method section and implementation section are not consistent. Sec. 2.4 defines the reliability weight ( Eq. 4 ) and a CE + Dice loss (Eq. 5) as the central self-training contribution. Sec. 3.1 instead describes a tri-partition score filter (discard if < 40, refine if 40–90, keep if ≥ 90), a fallback Dice check < 0.5, a Curvature Loss (λ = 0.01, undefined and uncited), and an “LLM-weighted Dice Loss” with weights = score/100.It explicitly states that “Rounds 1–3 … the model was fine-tuned … without LLM weighting” — which directly contradicts the claim that reliability weighting is essential. These two descriptions cannot both be the method that produced Table 1.Please reconcile. 3.Computational cost is not reported. Per image, the pipeline requires one LLM call plus three SAM forward passes, repeated over four adaptation rounds across the entire target training set. This is dramatically more expensive than the SFDA baselines and questionably practical at the very privacy-constrained clinical sites that motivate SFDA. No wall-clock, GPU memory, or LLM-API cost numbers are provided. 4.Limited evaluation scope. Evaluation is restricted to a single benchmark (M&Ms cardiac MRI). The conclusion claims generalizability to “more imaging modalities and clinical scenarios,” but no cross-modality or cross-organ evidence is presented. No statistical significance testing (paired tests across cases) is reported, despite modest test-set sizes per vendor. 5.Overclaimed framing. Terms such as “Cognitive Meta-Controller” and “cognitive reasoning” overstate what the LLM is actually doing — it is being used as a promptable visual inspector that outputs coordinates and a numeric score. This is useful, but the framing inflates the contribution.
- 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 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 paper proposes a creative direction — bringing a Multimodal LLM and a heterogeneous SAM ensemble into the SFDA loop — and reports large gains on M&Ms (+5.58% Dice, −50% ASSD over IPLC). However, several issues prevent acceptance in its current form. First, a key zero-shot baseline is missing: applying “LLM + Multi-SAM consensus” directly at inference without any target-model self-training. Without it, one cannot tell whether the SFDA loop actually contributes or whether inference-time prompting explains most of the gains. Second, the method section and implementation section are inconsistent. Sec. 2.4 defines a reliability-weighted CE + Dice loss (Eqs. 4–5) as the core self-training contribution, while Sec. 3.1 describes a tri-partition score filter, an undefined Curvature Loss, and an LLM-weighted Dice Loss, and explicitly states that Rounds 1–3 are trained “without LLM weighting” — contradicting the central claim. These two descriptions cannot both correspond to the method behind Table 1.Third, computational cost is not reported: one LLM call plus three SAM forward passes per image across four rounds is dramatically more expensive than the baselines, and questionably practical at the privacy-constrained sites that motivate SFDA. Fourth, evaluation is limited to a single benchmark (M&Ms cardiac MRI), with no cross-modality or cross-organ evidence and no statistical significance testing. Finally, the framing (“Cognitive Meta-Controller,” “cognitive reasoning”) overstates what the LLM is doing — it is used as a promptable visual inspector that outputs coordinates and a numeric score. I would reconsider my score if the authors convincingly address the first two points in the rebuttal.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
I still feel that the novelty of the paper is insufficient for publication at MICCAI.
Author Feedback
We thank all reviewers and the meta-reviewer for their constructive comments. We address the concerns in the order of the reviews.
R1.1 Visualization. We agree that qualitative comparison is important. We will add examples comparing source predictions, MLLM+SAM-refined pseudo-labels, and final predictions to show corrections such as missed RV, fragmented MYO, and boundary refinement.
R1.2 Computational cost. MLLM/SAM calls are an offline pseudo-label scoring/refinement cost, not a cost at every training forward pass. Target training samples are scored/refined first; during fine-tuning, the network only reads stored pseudo-labels and LLM-derived weights, so no MLLM/SAM call is invoked in each forward/backward pass. We will clarify this and add runtime/cost discussion.
R2.1 Visualization and statistics. We will add qualitative results as in R1.1, and report mean±std and paired statistical tests using per-case logs.
R2.2 SAM comparison and direct inference. Direct “LLM + Multi-SAM consensus” inference without target self-training gives mean Dice scores of 77.88/68.36/75.76 on Vendors B/C/D, averaging 74.00, while full RaMA achieves 85.68/82.13/80.58, averaging 82.80.Thus, the SAM/LLM inference module alone does not explain the gains; adaptation is essential.
R2.3 Ablation. Table 2 already shows the progressive effects of refinement and reliability weighting. We will revise the text to better separate source-only pseudo-labels, MLLM-guided refinement, reliability filtering/weighting, and iterative self-training.
R2.4 Necessity of the LLM. The LLM is not a final predictor; it acts as a semantic error checker that converts anatomical errors into FP/FN prompts. Simple uncertainty or boundary heuristics, e.g., IPLC, can sample points, but cannot judge organ-level errors such as missing RV or fragmented MYO.
R2.5 Reproducibility. We will add the thresholds 40/90, fallback Dice threshold 0.5, per-organ score weighting, training rounds, and learning rates.
R3.1 Direct inference. As shown in R2.2, direct LLM+Multi-SAM inference is clearly inferior to full RaMA, confirming that target-domain self-training, rather than test-time prompting alone, is essential.
R3.2 Method consistency. We apologize for the confusing description between Sec. 2.4 and Sec. 3.1.In implementation, reliability is realized in stages: (1) LLM scores triage pseudo-labels, discarding samples with any organ score <40, keeping high-confidence ones, and refining medium-confidence ones; (2) SAM3, MedSAM2, and SAM-Med2D generate refined masks by majority voting, with a fallback check to reject unstable refinements; (3) in initial target fine-tuning, normalized per-organ LLM scores weight the Dice loss together with curvature regularization; and (4) later self-training rounds regenerate pseudo-labels from the adapted model and use reduced learning rate without LLM-score weighting. Thus, “without LLM weighting” refers only to later rounds. We will revise Sec. 2.4/3.1 to separate filtering/refinement, score-weighted fine-tuning, and later self-training.
R3.3 Computational cost. As discussed in R1.2, MLLM/SAM operations are only used offline for pseudo-label scoring/refinement and are not invoked during each training forward/backward pass or deployment. We will make this explicit in the revised manuscript.
R3.4 Scope and wording. We agree that the current evaluation is limited to M&Ms cardiac MRI and will soften generalization claims. We will also replace overstrong wording such as “Cognitive Meta-Controller” with “MLLM-guided semantic error checker.”
Meta-review. The three requested points are addressed above: direct inference is substantially weaker than full RaMA, as shown in R2.2/R3.1; the reliability implementation will be clarified as a stage-wise procedure, as described in R3.2; and the LLM is justified as a semantic error checker rather than a standalone predictor, as explained in R2.4.
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.
This paper proposes RaMA, an innovative Source-Free Domain Adaptation (SFDA) framework that integrates a Multimodal LLM to generate spatial constraints for a multi-agent SAM ensemble (SAM3, MedSAM2, SAM-Med2D) to refine pseudo-labels. While the direction is creative, reviewers raised critical concerns regarding methodology consistency, baselines, and practicality.
-Fair performance of your “LLM + Multi-SAM consensus” directly at inference time without any target-domain self-training to prove the actual contribution of the SFDA loop (R3).
-Reconcile the explicit contradictions between the reliability-weighted loss defined in Section 2.4 and the tri-partition score filter/LLM-weighted Dice Loss described in Section 3.1 (R3).
-Justify the necessity of the LLM by comparing it against simpler heuristic prompting strategies.
- 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.
Following the author feedback, a clear majority of the review panel (2 Accept, 1 Reject) supports the publication of this work. The authors successfully vindicated the necessity of their self-training loop by providing the missing baseline comparisons and clarifying the stage-wise implementation of the reliability-weighted loss. The final manuscript must strictly include the promised statistical validations and qualitative segmentation figures to address the remaining minor critiques.
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.
After carefully considering the authors’ rebuttal and the reviewers’ final recommendations, the AC believes that the authors have satisfactorily addressed the majority of the concerns raised during the review process and that the paper meets the conference’s acceptance standards. While one reviewer ultimately questioned the novelty of the proposed method, it is worth noting that the same reviewer initially identified novelty as one of the paper’s primary strengths. This inconsistency weakens the basis of the novelty concern, particularly given the authors’ clarifications in the rebuttal. Overall, the AC finds that the paper presents a sufficiently original and well-evaluated contribution to warrant acceptance.