List of Papers Browse by Subject Areas Author List
Abstract
Test-Time Adaptation (TTA) has emerged as a strategy to adapt pretrained models to unseen data during inference without requiring retraining or access to large datasets. Despite recent progress, most existing TTA approaches rely on fixed adaptation strategies, lacking explicit mechanisms to determine whether a test sample requires adaptation and to what extent. As a consequence, adaptation may be unnecessary or suboptimal, especially in scenarios where test samples exhibit heterogeneous degrees of distribution shift, as in the case of medical imaging. Moreover, existing TTA methods are typically developed as by-design solutions that are tightly coupled to specific tasks. This limits modularity and hinders the development of adaptation frameworks that can be flexibly attached to pretrained models. To address these limitations, we propose a novel TTA framework based on a Multi-Agent System formulation that enables sample-specific adaptation and separates adaptation control from the pretrained task model. Our method decomposes TTA into four interacting agents responsible for task execution, distribution shift monitoring, adaptation selection, and model update. We validate our approach by carrying out extensive experiments on two medical image-to-image translation tasks, T1-to-T2 MRI translation and low-dose CT denoising, across four publicly available datasets. Results show consistent improvements over the state-of-the-art. Code is available at https://github.com/arco-group/Multi-Agent-TTA.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2376_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/arco-group/Multi-Agent-TTA
Link to the Dataset(s)
N/A
BibTex
@InProceedings{IelIre_Multiagent_MICCAI2026,
author = { Iele, Irene AND Di Feola, Francesco AND Soda, Paolo AND Sicilia, Rosa AND Tortora, Matteo},
title = { { Multi-agent Test-Time Adaptation for Robust Medical Image-to-Image Translation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16890},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper addresses selective test-time adaptation for medical image-to-image translation. The proposed decomposition into predictor, monitoring, selection, and adaptation agents is conceptually clean, and the evaluation spans two tasks and multiple datasets.
- 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 framework is modular and predictor-agnostic, and the experiments on both GAN-based and flow-matching-based predictors support that claim to some extent. The ablation study is useful in showing that monitoring, actuator selection, and sample selection all contribute to the final behavior.
The paper goes beyond a single benchmark and evaluates on both MRI translation and CT denoising, which is a meaningful strength.
- 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 serious issue is the calibration protocol for the selection threshold tau in the CT experiments. The paper states that tau is determined from the full available test set and then reused during evaluation on the abdominal and chest test sets. This means the evaluation data are used to define the adaptation policy itself. That is not a valid test-time protocol and materially compromises the reported CT results. Threshold calibration must be performed on a validation split or on source-domain held-out data only.
Reconstruction error is used as a proxy for distribution shift and as the basis for deciding whether adaptation is beneficial. That is plausible, but the paper does not directly evaluate whether this signal predicts adaptation benefit.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper’s experimental part (CT) has serious test-set leakage.
- 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
Review #2
- Please describe the contribution of the paper
The paper proposes a modular multi agent framework for test time adaptation in medical image to image translation. Rather than applying a single fixed adaptation strategy to all test cases, the method decomposes the process into four coordinated components. A predictor agent produces the translated image using a pretrained model. A monitoring agent estimates distribution shift through reconstruction error. A selection agent determines whether adaptation is necessary and chooses which feature levels should be adapted. An adaptation agent then applies lightweight feature level one by one convolutional transformations at test time. The method is evaluated on T1 to T2 MRI translation and low dose to high dose CT denoising across multiple datasets, using both GAN based and flow matching based predictors.
- 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.
- A key strength is the sample specific adaptation policy. Existing test time adaptation baselines often adapt every sample, whereas this method adapts only selected cases and still reports improvements in PSNR and SSIM. That is a practically meaningful idea under heterogeneous domain shift.
- The experimental section is reasonably broad. The paper includes MRI and CT tasks, several external datasets, and ablation studies examining the roles of monitoring, adaptation, and selection.
- 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 method assumes that test time domain shift can be corrected by channel wise feature remapping while preserving spatial structure. This is plausible for mild feature mismatch, but the paper does not provide enough information on how much the mismatch is allowed for their method to work.
- For samples selected for adaptation, the method evaluates many actuator configurations and performs multiple adaptation iterations for each one. It is unclear whether there is any early stopping rule, or whether all iterations are always executed once a configuration has been chosen.
- The inference cost also appears potentially substantial. Even if only a subset of test samples is adapted, evaluating many actuator subsets with repeated optimization steps may still introduce considerable overhead. The paper would be stronger with a clearer analysis of runtime relative to simpler test time adaptation baselines.
- 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.
(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 meaningful and reasonably novel framework with clear practical value. The paper goes beyond standard test-time adaptation by introducing a sample-specific, modular decision process rather than uniformly adapting every test case. This is a useful idea for medical image translation, where the amount and type of domain shift can vary substantially across samples.
The separation into predictor, monitoring, selection, and adaptation agents gives the framework a clear structure, and the experiments are also fairly broad, covering both MRI and CT tasks, multiple datasets, different predictors, and several ablations.
However, there are still open concerns about whether the proposed convolutional feature remapping is sufficiently expressive for more substantial domain shifts. In addition, when many actuator configurations and repeated optimization steps are evaluated, and the runtime tradeoff relative to simpler baselines is not fully characterized.
- 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.
The authors did not provide sufficient information or evidence to address the concerns raised about the paper. For example, in response to the review comment, ‘This is plausible for mild feature mismatch, but the paper does not provide enough information on how much mismatch is allowed for the method to work,’ the authors replied that ‘The evaluated MRI and CT shifts go beyond mild mismatch while preserving the spatial correspondence required by the translation task.’ However, this response does not provide quantitative evidence or a clear characterization of the mismatch tolerance range, and therefore does not adequately address the concern.
Regarding the concern about the inference cost of the proposed method compared to the baseline, the authors also did not provide sufficient quantitative information in the rebuttal.
Review #3
- Please describe the contribution of the paper
The paper introduces a multi-agent-based pipeline for test-time adaptation of pre-trained deep learning models. The authors propose adaptation on a per sample basis.
- 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 introduces a clear decomposition into monitoring, selection, adaptation, and prediction agents, which provides a structured and interpretable framework. 2)The method adapts only a subset of intermediate representations via actuator modules, which is more efficient and potentially more stable than full-model adaptation. 3) The use of reconstruction errors across multiple representation levels enables label-free detection of distribution shift, making the approach practical for real-world medical settings.
- 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 use of random search over actuator subsets is computationally expensive and lacks guarantees, raising concerns about scalability and optimality, especially as the number of feature levels increases. 2) The framework requires multiple forward passes and adaptation steps per test sample, which may limit applicability in time-sensitive or large-scale medical imaging scenarios. 3) Reconstruction error may not always correlate with harmful domain shift, potentially leading to unnecessary or suboptimal adaptations; stronger justification or validation is needed.
- 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 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
1) How was K=50 chosen, how sensitive is the method to the inclusion/ exclusion of certain features. Also specify n, i.e. for each subset how many features are sampled, how sensitive is the model to n?
- 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 method has various components that need to be optimized, stability of the method is unclear. Also, per sample optimization seems time consuming, reporting compute time per sample may be useful for end users.
- 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 rebuttal improves clarity regarding the adaptation protocol and actuator search strategy. While concerns regarding heuristic selection of hyperparameters and computational cost persist, I recommend “Accept” since the paper shows promising performance across datasets and holds practical merit.
Author Feedback
We thank the reviewers and meta-reviewer for their feedback. R1, CT threshold calibration: The manuscript sentence was incorrectly worded. The threshold tau is computed from the full source-domain Mayo abdominal training set, not from the CT test sets. Since the CT Predictor Agent is trained on Mayo abdominal scans, this is the natural reference for the upper tail of the in-distribution reconstruction error. Tau is fixed before evaluating the abdominal and chest test cohorts, and the same fixed threshold is then applied to both CT test sets. No evaluation sample is used to define the adaptation policy. The reported results were obtained under this protocol. The dynamic strategies, sigma2 and sigma3, are initialized from this reference and update tau online during inference. We will correct the manuscript accordingly. R1/R2/R4, Monitoring signal validity: Distribution shift is quantified through reconstruction errors produced by the Monitoring Agent, whose observers are trained on source-domain representations and frozen at test time. Higher output-level error indicates deviation from the source-domain distribution. This signal is a label-free, unsupervised shift score, not a deterministic predictor of adaptation gain. The evaluated shifts are not simulated perturbations, but arise from real distribution differences: cross-cohort, population, and pathology-driven shifts in MRI, as reported in Tab. 1, and anatomical-domain shifts in CT. Fig. 3 validates the monitoring-selection mechanism: among selected samples, the GAN-based Predictor Agent exceeds 80% win rate in 3/4 metric-task settings, reaching 98.6% in CT denoising. For the flow-matching Predictor Agent, PSNR improvements remain prevalent; SSIM shows more balanced win/loss in MRI, consistent with the stronger baseline and smaller improvement margin. These results indicate that the policy tends to trigger adaptation when beneficial. Sensitivity to synthetic shift severity is future work; here, we focus on real external-cohort and anatomical shifts. R2, Applicability of feature-level adaptation: The 1x1 actuators perform channel-wise feature remapping while preserving spatial structure. They address shifts in intensity, contrast, and representation statistics. The evaluated MRI and CT shifts go beyond mild mismatch while preserving the spatial correspondence required by the translation task. More extreme shifts requiring spatial transformations or large structural changes would require more expressive adaptation mechanisms, which could be explored within the proposed MAS framework. R2/R4, Computational cost, early stopping, and sensitivity to K/n: During TTA, the Predictor and Monitoring Agents remain frozen; only selected actuators are updated, limiting trainable parameters and instability risk. Adaptation is not applied to all samples. When the reconstruction error does not exceed tau, the sample follows the standard Predictor Agent inference path, with no actuator search or optimization. For selected samples, n is the number of intermediate representation levels exposed by the Predictor Agent and associated with candidate actuators. The Selection Agent samples K=50 actuator subsets uniformly at random; active actuators vary across subsets and are not fixed a priori. Fixing K bounds the search and avoids exhaustive enumeration of all actuator combinations. Each candidate subset is reset and optimized for at most T=10 iterations, with early stopping when the monitoring loss no longer improves. K=50 balances actuator-space exploration and test-time cost. Tab. 2 illustrates this selective design: sigma2 and sigma3 adapt fewer than 10% of MRI samples in all settings, compared to 100% for both competitors. This limits average test-time overhead for large-scale inference. R1/R2/R4, Reproducibility: We will better detail actuator sampling, early stopping, and frozen/trainable components in the revised manuscript, and release training, inference, and TTA code upon acceptance.
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.
Reviewer assessment is inconsistent. I see potential merits in the multi-agent TTA setup, which may be interesting for the community, maybe even beyond image to image translation. However, a few severe concerns were mentioned in the reviews. Please sort the weaknesses by importance and provide a response to the raised concerns in the rebuttal.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
Majority of reviewers vote for rejection of the paper. While the rebuttal improves the clarity regarding some methodological aspects, it still seems that core design choices and hyperparameters were not well justified. It also seems as if the contribution and novelty remain limited, therefore I agree with the majority assessment of reviewers.
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.
I recommend acceptance because:
-
The most critical concern raised by R1 (potential test-set leakage in the CT threshold calibration) was convincingly clarified in the rebuttal.
-
R4 recommended acceptance after the rebuttal.
The authors are expected to correct the manuscript wording that led to the calibration misunderstanding, provide a clearer runtime analysis, and better detail the sampling and early-stopping protocol in the camera-ready version.
-
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.
Though some concerns remain post-rebuttal, mainly related to experimental choices, this is a practical method that is overall evaluated broadly and could be useful beyond image-to-image translation. With TTA becoming increasingly popular, I believe that it is a valuable contribution to MICCAI.