List of Papers Browse by Subject Areas Author List
Abstract
Adapting foundation models to new clinical sites remains challenging in practice. Domain shift and scarce annotations must be handled by experts, yet many clinical groups do not have ready access to skilled AI engineers to tune adapter designs and training recipes. As a result, adaptation cycles can stretch from weeks to months, particu larly in few-shot settings. Existing PEFT methods either require manual adapter configuration or automated searches that are computationally infeasible in few-shot 3D settings. We propose SEA-PEFT (self-auditing parameter-efficient fine-tuning) to automate this process, moving towards realistic deployment in clinical environments. SEA-PEFT treats adapter configuration as an online allocation problem solved during fine-tuning rather than through manual, fixed-topology choices. SEA-PEFT uses a search–audit–allocate loop that trains active adapters, estimates each adapter’s Dice utility by momentarily toggling it off, and then reselects the active set under a parameter budget using a greedy knapsack allo cator. Exponential Moving Average and Interquartile Range smoothing, together with a Finite-State Ranking controller, stabilize the loop and improve reliability in high-noise few-shot regimes. On TotalSegmentator and FLARE’22, SEA-PEFT improves mean Dice by 2.4–2.8 points over the strongest fixed-topology PEFT baselines across 1/5/10-shot settings while training <1% of parameters. For reproducibility purposes, we made our code publicly available at https://github.com/tsly123/SEA_PEFT
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3785_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/tsly123/SEA_PEFT
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LySon_Selfauditing_MICCAI2026,
author = { Ly, Son Thai AND Nguyen, Hien V.},
title = { { Self-auditing Parameter-Efficient Fine-Tuning for Few-Shot 3D Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose a method for online search of optimal PEFT configurations for medical image segmentation in the context of few-shot finetuning for usage at a new site, as opposed to prior offline search methods, based on a search-audit-allocate loop. They show theoretical guarantees and fairly strong empirical results over suitable baselines.
- 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 submission is well-written and structured, and well-contextualized in the literature. Domain shift problems are remain a very real issue despite the rise of foundation models helping somewhat. Additionally, the challenge of new clinical sites not typically possessing an AI practitioner able to do the necessary adaptation and experimentation required to mitigate domain shift is a realistic concern. Altogether, this work tackles an important problem in the need for fully automated, guaranteed adaptation methods for medical image segmentation/analysis. 2.The theoretical guarantee results support the legitimacy of the proposed methods, which is frankly rare in MICCAI. The authors managed to fit this in despite length constraints and no supplementary while still having room for ample experiments. 3.The core idea is novel and interesting, and points out clear limitations in other adaptation/experimentation frameworks. The search-audit-allocate algorithm is conceptually sound in my opinion, and implemented in a reasonable fashion. 4.The experimental results are strong, comparing a good range of adapters on suitable base models, on standard, fairly challenging multi-organ segmentation baselines. The performance gains over the baselines are noticeable on TotalSeg, and somewhat for FLARE. The compute time is also very reasonable for a single GPU (Table 3), especially since everything is automated, and this is a one-time process when adapting a model to a new site.
- 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.As mentioned, performance gains are inconsistent and small on average for FLARE (Table 1). Can the authors explain why this is the case, compared to TotalSeg, and if there is a corresponding limitation of their approach? Could it be related to FLARE being multi-class vs. binary? Could the search space be harder to elucidate for more complex tasks?
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
There is a slight typo in Fig 2 legend: SA-PEFT instead of SEA-PEFT.
- 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper proposes a novel, interesting, principled method to attack the key problem of domain shift and adaptation for medical image segmentation. The authors show notable performance gains compared to fair state-of-the-art baselines on two wide-ranging multi-organ-segmentation datasets, although the gains on FLARE for multi-class segmentation are smaller. Notably, the manuscript also provides theoretical results/guarantees, unusual for MICCAI, while simultaneously seeming suitable for realistic clinical use by non-AI experts. Altogether this is a high-quality submission that will likely be of interest to a wide audience at MICCAI, and I recommend its acceptance.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors sufficiently addressed my only concern, and I think they did well to address the concerns from the other reviewers and the metareviewer. I support acceptance.
Review #2
- Please describe the contribution of the paper
SEA-PEFT is a framework that automates adapter configuration for parameter-efficient fine-tuning of 3D medical image segmentation foundation models in few-shot clinical settings. The core insight is to treat adapter selection — which adapter type (LoRA, AdaptFormer, Affine-LN), which topology (Serial, Parallel, or Composite), which rank, and which layer — as an online resource allocation problem solved during a single training run rather than through expensive offline search. This is achieved through a repeated search–audit–allocate loop: active adapters are trained for several steps, then each adapter’s marginal Dice utility per parameter cost is estimated by temporarily toggling it off on a validation set, and a greedy knapsack allocator activates the best-value combination within a fixed parameter budget. EMA and IQR smoothing stabilize noisy utility estimates in low-data regimes, and a Finite-State Machine prevents configuration chatter by requiring several consecutive consistent allocation proposals before committing any change. Once the loop converges, the optimal adapter set is selected and the model is retrained from scratch with only those adapters active. On TotalSegmentator and FLARE’22 benchmarks across 1, 5, and 10-shot regimes, SEA-PEFT consistently outperforms all fixed PEFT baselines without any manual configuration.
- 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 problem is well-motivated. Fine-tuning foundational models with limited amount of labeled data is a common strategy for medical image segmentation. However, it is not trivial to find the best PEFT method for fine-tuning and requires expertise which creates a bottlenect in clinical deployment. Automating this within a single training run directly addresses a meaningful gap.
2- The idea is interesting and intuitive. The on/off perturbation-based utility estimation is simple, intuitive, and model-agnostic, and directly measures each adapter’s marginal Dice contribution. Additionally, the paper proposes a computationally efficient way of estimating adapter utility within a single training run by auditing only a small mini-batch of adapters per cycle, making the computational cost scale with data size rather than configuration space size. Unlike the common way of fine tuning foundational models of selecting a fixed adapter for every layer, SEA-PEFT can discover that different layers benefit from different adapter types.
3- Strong experimental results. SEA-PEFT achieves the highest mean Dice across all few-shot regimes (1, 5, 10), both datasets (TotalSegmentator and FLARE’22), and both backbone families (FSEFT and SuPreM), with improvements of 2.4–2.8 Dice points over the strongest fixed-topology baseline. The consistency of gains across diverse settings strengthens the credibility of the results.
4- Practical efficiency. Completing full adaptation in 2.5–6.5 hours on a single V100 GPU while training only 0.2% of parameters makes SEA-PEFT realistically deployable in clinical environments with limited computational resources, which is consistent with the paper’s motivating scenario.
- 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 selected adapter configurations are not reported. The paper’s central claim is that automated configuration selection adapts to each dataset and shot regime. Yet S^*_{final} is never explicitly shown for any experiment — there is no table or figure revealing which layers got adapters, which topology was preferred, or which ranks were selected on TotalSegmentator vs. FLARE’22.This would be an interesting analysis to understand whether the method chooses very different adapters for different datasets.
2.The paper reports results averaged over three random seeds but omits standard deviations or confidence intervals from all result tables. Reporting only mean Dice scores is insufficient to understand the performance of the method with different seeds. Additionally, it is unclear whether 3 seeds were used when selecting adapters with SEA-PEFT or during the final training once the final adapters are set.
3.The validation set construction in few-shot settings is never clarified.The entire audit mechanism relies on a labeled validation set D_val separate from the training shots. In the 1-shot regime this is particularly problematic — if the single labeled scan is split between training and validation, training quality degrades significantly. If a separate labeled scan is reserved for validation, the true annotation budget is effectively doubled. The paper never states how D_val is constructed, which is a critical missing experimental detail.
4.No ablation studies on any component. There is no ablation isolating the contribution of EMA smoothing, IQR penalization and the FSM stabilizer. The contribution of these components is crucial since they are claimed as one of the main contributions of the paper.
5.The “no manual configuration” claim is overstated.SEA-PEFT still requires setting many parameters and the entire audit space design including which topologies and rank values to include. These are presented as inherited defaults from prior work, but a clinical user who is “not an AI engineer” would still need guidance on these choices, and no sensitivity analysis is provided to justify that they transfer reliably across new datasets or modalities.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 addresses a genuinely important and well-motivated problem — automating PEFT configuration for few-shot 3D medical image segmentation — and the core idea of on/off perturbation-based utility estimation is intuitive, elegant, and computationally practical. The experimental results are consistent and encouraging across multiple datasets, backbones, and shot regimes, and the practical efficiency of the method aligns well with the clinical deployment scenario the paper targets. There are couple of weaknesses I see as mentioned in the weaknesses section. I think the most critical ones among all are the followings:
1- Missing ablation studies (W4). EMA smoothing, IQR penalization, and the FSM stabilizer are all presented as key contributions of the paper, yet none of them are individually validated. Without ablations, it is impossible to know whether these components are genuinely necessary or whether a much simpler design — for example plain moving average smoothing without IQR, or no FSM at all — would perform comparably.
2- The validation set construction is never clarified (W3). This is particularly damaging in the 1-shot regime, which is one of the paper’s headline experimental settings. If a separate labeled scan is used for validation, the effective annotation budget is doubled, making the 1-shot claim misleading. If the single training scan is split, training quality is compromised. Either way the experimental setup needs to be explicitly described — as it stands, the 1-shot results cannot be properly interpreted.
In general I am positive about the paper. However, I think the weaknesses are major enough to suggest rejection. My initial rating is weak reject but I would be happy to increase my score after 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.
The rebuttal addresses some of my concerns, especially by clarifying the validation set construction and providing high-level ablation results for EMA/IQR and FSM. I also understand that the authors had limited space to respond to all reviewers’ comments, and I appreciate their efforts. However, I still find the response insufficient to raise my score because several critical points remain unresolved. First, the authors do not provide the actual standard deviations, confidence intervals, or per-seed results in the rebuttal, so the robustness of the reported gains remains unclear. This is particularly important because the improvements over the strongest baselines are modest in some settings. Second, the selected adapter configurations are only described qualitatively, without concrete per-dataset, per-shot, or per-layer details, which weakens the central claim that SEA-PEFT discovers meaningful dataset-specific configurations. Finally, although the authors clarify that D_{val} is a fixed public validation partition and is used consistently across methods, I remain concerned that excluding this labeled validation set from the annotation budget may make the “1-shot” setting less realistic from a clinical few-shot adaptation perspective. I acknowledge that this may be less problematic if the same validation protocol is applied fairly to all methods. Overall, while the rebuttal improves clarity, I still consider these remaining issues substantial and suggest reject.
Review #3
- Please describe the contribution of the paper
The authors formulate a 3D medical image few-shot PEFT configuration as an online resource allocation problem, bypassing computationally prohibitive offline architecture searches. It introduces an “on/off perturbation” auditing mechanism coupled with EMA+IQR smoothing and a Finite-State Machine (FSM) to dynamically select optimal adapter topologies during training.
- 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 paper focuses on an important and practical topic. 2.The paper is generally well-structured and relatively easy to understand and follow.
- 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 core “on/off perturbation” scoring relies on a strong, unverified assumption of independent adapter utility. This fundamentally ignores potential synergistic or antagonistic effects between different layers and adapter types, risking severely sub-optimal greedy allocations. 2.The omission of current automated or dynamic PEFT baselines (e.g., NOAH, AutoPEFT, AdaLoRA) is unacceptable. Dismissing these comparisons due to 3D computational costs without providing even a proxy 2D benchmark leaves the claimed optimality completely unverified. Moreover, AdaLoRA is not constrained by few-shot settings, and the authors themselves mention it in the text; why was it not included in the comparison? 3.Crucial empirical ablations for the heavily emphasized EMA+IQR tracker and FSM stabilizer are entirely missing. Readers cannot isolate or verify the actual performance impact of these specific engineering components versus the overarching framework. 4.The framework’s heavy reliance on a robust validation set D_val for high-frequency utility auditing contradicts the premise of a strict few-shot clinical setting. The authors fail to define the size of D_val or address the severe risk of architectural overfitting when validation data is critically scarce.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 primary reasons for recommending a weak rejection are the complete absence of comparisons against state-of-the-art dynamic or search-based baselines, an unverified assumption regarding the independent utility of adapters, and a lack of necessary ablation studies for the proposed stabilization mechanisms.
- 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 has addressed most of my concerns. Although AdaLoRA still achieves performance comparable to the proposed method, which was my main remaining concern, I am convinced by the authors’ response and would like to update my recommendation to accept.
Author Feedback
We thank all reviewers for their careful feedback. Per MICCAI 2026 guidelines, PDF supplementary is prohibited, so all requested ablations and clarifications are included inline below and will be incorporated in the camera-ready.
[Fair comparisons and Meta-Q2] All baseline numbers follow published configurations of [1] (MICCAI 2025) and [18] (MICCAI 2023). SEA-PEFT uses the identical backbone, D_train, D_val, D_test, random seeds, training protocol, and matched parameter budget across all experiments. D_val is the fixed public validation partition, entirely separate from the K training shots.
In the 1-shot setting, 1 labeled volume is used for gradient updates; D_val is not counted toward the few-shot budget (annotation cost is exactly K volumes). Three random seeds were used for both the search phase and the final retraining; results are on the held-out test set, never D_val. Mean+-std will be included in the camera-ready.
[Meta-Q1] Why NOAH/AutoPEFT are infeasible here. Both [32] and [33] require one complete end-to-end training run per candidate. Our audit space covers 120+ configurations; at 1-shot, one run takes 2.5h on V100 (Table 3), so a full sweep exceeds 300 GPU-hours per organ per shot setting, prohibitive for the per-site adaptation scenario we target, whereas SEA-PEFT explores the same space in one run.
AdaLoRA (new comparison; LoRA-only audit space, ranks {2,4,8,16,32}; identical protocol). 5-shot: SEA-PEFT avg. 77.60 vs. AdaLoRA 75.57 (Gall: 81.03 vs. 78.11; Eso: 48.62 vs. 43.76). 10-shot: 79.84 vs. 76.57 (Gall: 84.58 vs. 80.01; Eso: 52.46 vs. 45.72). SEA-PEFT wins by 2.0-3.3 Dice points because direct perturbation-based utility is more reliable than SVD-based saliency when gradient signals are noisy at 5-10 shots.
[Meta-Q3 and R#2-W5] Hyperparameters are fixed across all experiments, organs, datasets, and backbones (no per-task tuning): K=200, M=0.2N, beta=0.95, lambda_S=0.5, tau_act=tau_rank=3, Pmax=1%.
We swept K={100,500, 800,1000}, M={0.1N,0.5N,0.8N,1.0N}, and beta={0.90,0.925,0.975,0.999} individually. On TotalSeg (1/5-shot): mean absolute Delta Dice change relative to the default across all conditions is 0.12 (K), 0.64 (M), 0.31 (beta); worst-case degradation is -0.82, -0.55, -0.59.On FLARE (1/5/10-shot): 0.12, 0.14, 0.24 mean; -0.54, -0.53, -0.47 worst. Performance is bounded and stable across all sweeps without per-task tuning. We will revise “engineer-free” to “toward engineer-free” throughout.
[R#2-W1] Selected configurations. SA dominates across layer depths (PA secondary; SAPA selective); LoRA leads unit-type choice (AdaptFormer next; Affine-LN sparse). Allocations are layer-non-uniform and dataset-dependent, confirming the search discovers structured, dataset-specific configurations rather than a fixed pattern. Per-dataset breakdown will be added in the camera-ready.
[R2-W4 and R4-W3] Ablation results (DeltaDice vs. no component, TotalSeg).
EMA+IQR: avg. +2.83 (1-shot), +3.60 (5-shot) across 9 organs; largest gains on Aor +12.08 (1-shot) and +9.88 (5-shot), Gall +5.47 and +9.93.
FSM: avg. +2.15 (1-shot), +1.74 (5-shot), +1.58 (10-shot); largest gains on Gall +5.29, Eso +4.60, Sto +3.31 (1-shot). Both components are individually necessary. Full tables will be added in the camera-ready.
[R#4-W1] Adapter utility is not assumed independent. The on/off audit measures each adapter’s marginal Dice conditioned on the current active set, capturing co-adapter synergy and antagonism. The greedy allocator treats these as additive, bounded by the epsilon term in Eq. 8 and refined by repeated audit cycles. SEA-PEFT outperforming AdaLoRA across every organ and shot setting confirms the approximation is tight in practice.
[R#1-W1] FLARE gains are smaller because joint multi-class (9 organs) Dice is a noisier per-adapter signal than binary per-organ Dice, inflating the epsilon term in Eq. 8.The same pattern appears in [1] and [18]. We will add this discussion.
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 paper received mixed reviews, specifically one strong accept and two weak reject, one of whom remains positive and open to increasing the score. During the rebuttal, I would be especially interested in the authors’ feedback on the following points:
(1) The omission of current automated or dynamic PEFT baselines in the quantitative evaluation. In particular, please clarify what is meant by “not feasible for high-resolution 3D segmentation at 1–10 shots,” for example, in terms of compute time, GPU usage, or parameters compared to the results in Table 3.Please also elaborate on the exclusion of AdaLoRA from the comparisons. (2) Clarify what is used as the validation set in the few-shot regime. (3) Regarding the missing ablation studies, while new experiments are not allowed during rebuttal, these components involve a few hyperparameters. Please clarify whether they are fixed across all experiments or per task, and in the latter case, how this is handled in few-shot regimes.
Please also address other important concerns raised by the reviewers.
- 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 the rebuttal, the paper received 2 “Accept” and 1 “Reject” recommendations. One of the initial reviewers increased the initial score.
The rebuttal successfully clarified the exclusion of dynamic PEFT baseline methods due to their computational cost and provided positive comparisons with AdaLoRA. Also, the additional high-level results on the stability of performance under hyperparameter modifications are encouraging. The remaining concerns from R2 to recommend rejection point out: (1) Not providing standard deviation in performance; (2) Missing detailed analysis on selected configuration; and (3) The consideration of the validation set on the few-shot budget. These are all valid comments.
For (1-2), in my opinion, these comments are not invalidating and might be justified by space constraints. For (1), although performance gains may be tight in a specific scenario, the proposed SEA-PEFT procedure is motivated not only by performance but also by the practical utility of automatically selecting the PEFT configuration in a stable way across data regimes, backbones, and datasets.
Regarding (3), I find this concern more problematic. The validation data is used to guide the audit phase by computing the Dice score. Therefore, as R2 argues, it requires annotation and should, accordingly, be considered part of the annotation budget. At least, it should be acknowledged that a few-shot validation set is required for SEA-PEFT. According to the repository, the authors appear to use K=2 for validation. Results with SEA-PEFT (1 train shot + 2 validation shots) are positive compared to non-dynamic baselines using 5 shots (e.g., in Table 1). Therefore, results suggest an efficient profit from such validation data. The authors should revise the manuscript to make the validation set design more explicit.
Based on the above, the remaining concerns do not invalidate the manuscript’s strengths, which cover a timely, practical topic on the more automated adaptation of medical foundation models and propose an original adaptation procedure with positive empirical validation across backbones and datasets. I recommend acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper introduces SEA-PEFT, an online allocation framework designed to automate parameter-efficient fine-tuning (PEFT) configurations for few-shot 3D medical image segmentation. By utilizing a search-audit-allocate loop during training, the framework estimates task-level adapter utility through model-agnostic on/off perturbations, bypassing the need for computationally prohibitive offline architectural searches. The paper received split post-rebuttal recommendations consisting of one Strong Accept, one Accept, and one Reject. Reviewers highly praised the clinical practicality of automating site-specific adaptation without dedicated AI engineers, the conceptual soundness of the online knapsack formulation, and the rare inclusion of theoretical convergence bounds within length constraints. Furthermore, extensive multi-organ evaluation on TotalSegmentator and FLARE datasets demonstrates noticeable gains of 2.4–2.8 mean Dice points over competitive fixed-topology baselines within highly reasonable GPU execution timeframes.
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.
Good paper with interesting idea on search– audit–allocate loop for PEFT in 3D fewshot segmentation. The rebuttal resolves most of the concerns and therefore I recommend acceptance.