List of Papers Browse by Subject Areas Author List
Abstract
Deploying medical AI across resource-constrained institutions demands data-efficient learning pipelines that respect patient privacy. Federated Learning (FL) enables collaborative medical AI without centralizing data. Yet, real-world clinical pools are inherently open-set: beyond the in-distribution (ID) target classes, the unlabeled pool also holds out-of-distribution (OOD), non-target samples such as imaging artifacts, background tissue, and other pathologies. Standard Active Learning (AL) query strategies mistake this noise for informative samples, wasting scarce annotation budgets. We propose PromptGate, a dynamic VLM-gated framework for Open-Set Federated AL that purifies unlabeled pools before querying. PromptGate introduces a federated Class-Specific Context Optimization: lightweight, learnable prompt vectors that adapt a frozen BiomedCLIP backbone to local clinical domains and aggregate globally via FedAvg—without sharing patient data. As new annotations arrive, prompts progressively sharpen the ID/OOD boundary, turning the VLM into a dynamic gatekeeper that is strategy-agnostic: a plug-and-play pre-selection module enhancing any downstream AL strategy. Experiments on distributed dermatology and breast imaging benchmarks, with a random acquisition rule, show that while static VLM prompting degrades to ∼60% ID purity, PromptGate maintains >95% purity with 98% OOD recall. The source code is available at http://github.com/albarqounilab/promptgate.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5058_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/albarqounilab/promptgate
Link to the Dataset(s)
N/A
BibTex
@InProceedings{NesAde_PromptGate_MICCAI2026,
author = { Nesturi, Adea AND Gaviria, David Dueñas AND Zeng, Jiajun AND Albarqouni, Shadi},
title = { { PromptGate: Client-Adaptive Vision–Language Gating for Open-Set Federated Active Learning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16879},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
PromptGate is a lightweight VLM-gated module for open-set federated active learning. The method is validated on two different 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.
1.The combination of clinical problem and technical solution is working well in this paper. 2.Open-set learning problems are under-studied in the MICCAI setting. This paper has the potential to generate some serious conversations. 3.The experimental design is strong and the comparison is thorough - I liked it.
- 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.While the formal definition of open-set in section 2.1 is correct, the authors used an ambiguous definition in abstract. Please fix “yet real-world clinical pools are inherently openset, containing out-of-distribution (OOD) noise such as imaging artifacts and wrong modalities. “ 2.Method description at the end of introduction lacks clarity. Please resolve it.
- 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.
(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?
An open-set learning is novel enough for MICCAI audiences. The experiment is also solidly complemented the method.
- 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.
I stick to my original decision.
Review #2
- Please describe the contribution of the paper
This paper tackles the challenge of open-set federated active learning (OS-FAL) by introducing PromptGate, a vision-language gating mechanism. PromptGate is specifically designed to filter out out-of-distribution (OOD) samples before the active learning phase. The authors validate the efficacy of PromptGate by comparing it against three filtering methods across two distinct federated 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.
1.The proposed PromptGate serves as an acquisition-agnostic, plug-and-play gating module. By filtering out OOD samples before the active learning phase, it can be integrated with various active learning strategies. 2.The paper is well-organized and easy to read.
- 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 technical novelty of PromptGate is limited, as dual-prompt designs are widely adopted in federated learning [R1-R2]. Furthermore, since FOCoOp [R3] provides a more comprehensive approach for ID/OOD separation using ID global, ID local, and OOD prompts, the novelty of PromptGate over FOCoOp remains unclear. 2.While the authors highlight the mixed prompt design in Sec. 2.1, their ablation study reveals that using solely local prompts actually outperforms it, which raises concerns about the necessity and effectiveness of the mixed strategy. 3.On the FedEMBED dataset, Coldstart and Baseline yield lower purity but competitive or even better BMA compared to PromptGate. This raises concerns about the necessity of filtering out all OOD samples, as the experimental results indicate that a small amount of OOD samples does not negatively impact model optimization.
[R1] FedPHA: Federated Prompt Learning for Heterogeneous Client Adaptation. ICML 2025.[R2] Global and Local Prompts Cooperation via Optimal Transport for Federated Learning. CVPR 2024.[R3] FOCoOp: Enhancing Out-of-Distribution Robustness in Federated Prompt Learning for Vision-Language Models. ICML 2025.
- 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.
(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?
Considering the limited technical novelty and incomprehensive experimental validation, I recommend rejecting the paper in its current form.
- 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 #3
- Please describe the contribution of the paper
The paper proposes PromptGate, a dynamic VLM-gated framework for Open-Set Federated AL. It uses a frozen VLM with federated prompt tuning to filter OOD samples at each client before active learning query. Experiments show PromptGate effectively converts the Open-Set problem into Closed-Set problem.
- 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 work is well motivated, as it brings together open-set active learning and federated active learning in a medical imaging setting, where both privacy constraints and data quality issues are highly relevant. In practice, active learning can easily mistake OOD samples for highly informative ones, leading to wasted annotation budget. Addressing this problem is therefore both meaningful and practically important.
The proposed method is lightweight, introducing only a small number of additional parameters, while effectively narrowing the performance gap between open-set and closed-set settings. The gains on intermediate metrics are particularly clear, suggesting that the method is effective at improving sample selection quality.
The global/local prompt decomposition is also a reasonably novel design in the federated setting. The global prompts are intended to capture shared semantic knowledge across clients, while the local prompts adapt to client-specific OOD characteristics. This design is well aligned with the realistic heterogeneity across hospitals and medical centers, where scanner types, acquisition protocols, and artifact distributions often differ substantially.
- 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.
In Sec 2.2, the paper first defines p_c^k(x) and \hat{y}^k(x), stating that “PromptGate filters U^(r)_k into an ID-candicate pool”. Immediately after, it defines p_c^k(x) and \hat{y}^k(x) again and repeats very similar line “PromptGate filters U^(r)_k into an ID-candicate pool(VLM-gated)”. it’s repetitive and confusing.
Insufficient analysis of early-round instability: 1.The manuscript notes that query precision (QP) can dip at round 2 after the first CoOp update on a small 128-shot subset, but provides no analysis of how severe or common this is across clients. 2.There is no experiments show whether this early instability could lead to the irreversible loss of informative ID samples that are incorrectly gated out, potentially harming final model performance even though it’s mentioned in Sec 2.2 that“empirically, ID leakage into \epsilon_k^(r) is negligible.”.
The description of the gating mechanism is also somewhat unclear. Although the paper refers to a “high-confidence ID candidate pool” and defines p_max(x), the actual formulation only keeps samples whose argmax prediction belongs to an ID class, without using an explicit confidence threshold. It is therefore unclear whether the gate is category-based or confidence-based.
Some implementation details should also be clarified, such as whether prompt learning is retrained at each round or updated incrementally, and how the temperature parameter is selected.
Finally, although the paper highlights the mixed global+local prompt design, the results suggest that the Local variant is often the strongest or most stable. This raises the question of whether the main benefit comes from local personalization rather than federated global aggregation.
More complete statistics and ablations, such as different OOD ratios, initial labeled-set sizes, prompt token numbers, and per-client performance, 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 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?
Overall, this paper studies a meaningful and practical problem at the intersection of open-set active learning and federated learning for medical imaging. The proposed PromptGate is lightweight, reasonably well motivated, and shows clear improvements in query purity, with a sensible global/local prompt design for handling cross-client heterogeneity. The idea is interesting and the intermediate results are promising.
The most compelling evidence in the paper so far focuses on intermediate metrics like purity, while the improvement in the performance of the final downstream task is not yet stable, significant, or sufficient. Meanwhile, some key aspects of the method description remain unclear, such as whether the gate is based on class or confidence, the training and update method of the prompt, and whether the independent effect of global aggregation truly holds. In addition, the statistical information and ablation experiments are not yet complete. I believe this work is on the right track and the method has potential, but at this stage, the chain of evidence is insufficient to support a stronger acceptance.
- Reviewer confidence
Somewhat confident (2)
- [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.
N/A
Author Feedback
We sincerely thank the AC, R1, R2, and R3.All reviewers find the work “well motivated”, the “experimental design strong and thorough” (R1), “clinical problem and technical solution working well” (R1), and “serves as an acquisition-agnostic, plug-and-play gating module” (R2).
[R2, AC] Novelty over FOCoOp, FedOTP, and FedPHA. We are glad this concurrent work validates our global/local prompt decomposition, but all three address closed-set federated classification on fixed labeled datasets. FedOTP uses UOT over patches; FedPHA uses SVD projection; FOCoOp trains OOD pseudo-prompts with post-hoc MCM scoring — none has an acquisition gate, an unlabeled pool, or a mechanism to consume the supervised non-target label OS-FAL provides. PromptGate contributes: (a) the first learnable-prompt VLM gate for OS-FAL (CSC CoOp, global/local decomposition, (K+1)-class cross-entropy); (b) decoupling gate from acquisition rule; (c) converting OS-FAL into approximately closed-set AL across 6 strategies x 2 datasets. Prompts are warm-started and refined each AL round from oracle labels, progressively sharpening the ID/OOD boundary — driven by the AL loop itself, absent from all three papers. We will add them to the related work.
[R2, R3, AC] Local outperforms Mixed, is global aggregation necessary? No single configuration dominates universally. Table 1: Local leads on Random, FEAL, and LfOSA on FedISIC; Global leads on OpenPath (86.6 purity, 64.4 BMA) and FEAL purity; FedEMBED follows the same mixed pattern. The right choice depends on OOD variability across clients: when OOD artifacts are similar across sites, global prompts capture a stronger shared prior; when OOD is client-specific (as in FedISIC), local adaptation is more beneficial. We offer all three as dataset-dependent options; global aggregation always preserves privacy-preserving collaboration and distributed annotation cost across institutions that cannot share patient data.
[R2] Necessity of filtering on FedEMBED. Two factors explain the small BMA gaps: (1) OOD prevalence is low (4.8-13.2%), so unfiltered AL is already mostly ID by chance; (2) FedEMBED OOD labels are pseudo-labels. The decisive test is FedISIC (50% OOD): PromptGate yields +22% purity and +0-6% BMA across all 6 strategies. In short, PromptGate is safe under low-OOD and substantially helpful under high-OOD, precisely what clinical deployment requires.
[R3] Gating mechanism. The gate is category-based over K+1 classes: x is retained iff y-hat^k(x) in {1,…,K}. We avoided an explicit confidence threshold to prevent ambiguity with Entropy-based AL scores; implicit confidence filtering is achieved via the temperature-scaled softmax — we will clarify this in the text. [R3] Early-round instability in FedISIC and ID leakage. The round-2 QP dip has two compounding causes. First, at R=1 the VLM has seen only 128 shots and remains uncertain about OOD; uncertainty-based strategies (Entropy, FEAL, PAL, LfOSA) select ambiguous samples, disproportionately OOD, dragging QP down. This holds for the baseline (Static VLM as well). Second, reported QP values are means across 3 seeds x 6 AL methods, so uncertainty-based methods dominate (Entropy: 11.0%, FEAL: 11.6%, PAL: 14.1%, LfOSA: 15.8%; Random: 61% OpenPath: 97.9%). From R=3 onward, as the VLM accumulates oracle labels, this vanishes, and QP recovers to near-perfect. Samples in epsilon_k^(r) are not permanently removed: they remain in U_k and are re-evaluated each round. On FedISIC, all variants stabilise at ≤ 1.3% leakage, meaning ≥ 98.7% of true-ID samples are correctly retained in the ID pool, replacing the informal “negligible” claim.
[R3] Implementation details. Prompts are updated incrementally each round (15 epochs SGD, warm-started). tau_VLM = 0.0117 follows BiomedCLIP convention. Details will be added to Sec. 2.3.[R1, R3] Writing clarity. We will: a) align the abstract open-set phrasing with Sec. 2.1; b) rewrite Sec. 1; c) remove the duplicate definitions in Sec 2.2
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 authors are advised to address all the concerns raised by the reviewers, especially: (i) discuss the novelty; (ii) design choices such as prompt mixing and discarding OOD samples; and (iii) clarify the instability in the early rounds.
- 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.
While all the reviewers appreciated the presentation and the evaluations of the paper, they received mixed ratings (2.Accept and 1.Reject), and mostly the disagreement was on the novelty of the method. After going through the reviews and the rebuttal from the authors, I align with the fact that federated learning in an open-set setting can bring interest to the wider community.
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 main concerns arising from the initial reviews were related to novelty, prompt mixing choices, and early-round instability. All these issues have been addressed adequately in the rebuttal. The pending concern is that the use of local and global prompts is not novel. While this may be true, the application of this concept to Open-Set Federated Active Learning is sufficiently novel. Hence, the paper can be accepted.
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.
Despite moderate concerns about novelty, the rebuttal effectively clarifies the contribution: PromptGate is the first learnable VLM-gated module tailored for OS-FAL, with clear distinctions from concurrent works (e.g., FOCoOp, FedPHA). Two reviewers (R1 and R3) shifted to or maintained acceptance after the rebuttal, while the single reject opinion does not outweigh the overall positive assessment.