List of Papers Browse by Subject Areas Author List
Abstract
nnU-Net is the de facto standard for medical image segmentation, yet its architectures remain computationally expensive. We revisit nnU-Net self-configuration from an optimization angle: instead of heuristic architecture design or post-hoc pruning, we reformulate nnU-Net training as a proximal optimization problem with non-smooth regularizations on network structures. We design the reguarlization as a mixed L1 and group L2 loss. Using proximal optimization for this mixed loss enables safe removal of entire channels without harming the segmentation loss. It unifies weight and structure learning in a single, integral training process; and ``slim” nnU-Net architectures emerge as the result of optimization. We carry out experiments on both 2D cardiac MRI (ACDC) and 3D brain tumor segmentation (BraTS 2023) datasets, and benchmarked our method against a range of established baselines. We show that for both 2D and 3D nnU-Net, our method removes >99% of the weights structurally, while preserving near-baseline Dice performance. We further show that the reconfigured nnU-Nets deliver substantial run-time gains, with up to 96% FLOP reduction for 2D and 86% for 3D. Compared to baselines, slim nnU-Net achieves the best accuracy–efficiency trade-off. Our method enables more efficient inference and reduced hardware requirements for broad applications of nnU-Net.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4849_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhaYid_Slim_MICCAI2026,
author = { Zhao, Yidong AND Zhang, Yi AND Tao, Qian},
title = { { Slim nnU-Net: Revisiting nnU-Net Self Configuration from an Optimization Angle } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper revisits nnU-Net self-configuration from an optimization perspective rather than heuristic architecture design. It formulates nnU-Net training as a proximal optimization problem and introduces a sparse group lasso regularization, enabling the removal of entire channels without degrading segmentation performance. The authors evaluate their method on two datasets, ACDC and BraTS 2023, and demonstrate that it can remove more than 99% of the weights while preserving near-baseline Dice performance.
- 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.
- Revisiting nnU-Net self-configuration from an optimization perspective is novel, given that nnU-Net is a widely used segmentation model.
- Using proximal optimization for a mixed loss enables the removal of entire channels without degrading segmentation performance.
- The experiments in the paper are comprehensive.
- 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 authors neither provide a link to the source code nor indicate whether it will be released, which affects reproducibility.
- 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.
(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?
Mainly because revisiting nnU-Net self-configuration from an optimization perspective is novel, given that nnU-Net is a widely used segmentation model.
- Reviewer confidence
Not confident (1)
- [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 contribution is good and nnU-Net might benefit from this paper.
Review #2
- Please describe the contribution of the paper
This paper revisits nnU-Net self-configuration from an optimization perspective. Instead of using heuristic pruning or post-hoc architecture redesign, it formulates nnU-Net training as a proximal optimization problem with mixed element-wise ℓ1 and group-wise ℓ2 regularization. This encourages both weight sparsity and structured channel sparsity during a single training process, allowing entire channels to be removed after training. The method is evaluated on 2D cardiac MRI segmentation (ACDC) and 3D brain tumor segmentation (BraTS 2023), where it reportedly achieves extreme compression with strong accuracy-efficiency tradeoffs relative to several pruning 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.
A major strength is that the paper tackles a very relevant practical problem: nnU-Net is effective but often computationally heavy, especially in 3D. The motivation is clear and important for broader deployment. The paper frames this not as generic pruning, but as an extension of nnU-Net’s self-configuration philosophy toward optimization-based structure learning, which is a clean and appealing idea. The formulation is simple and principled: sparse-group regularization is introduced directly into the training objective, and proximal updates are used because the regularizer is non-smooth. The derivation in the methods section is easy to follow, and Figure 1 on page 3 clearly illustrates how sparse weights and channels emerge and lead to a compact slimmed model.
The paper does more than report parameter counts: it shows Dice-vs-FLOP tradeoffs in Figure 2, optimization dynamics in Figure 3, layer-wise retention patterns in Figure 4, and actual CPU/GPU timing in Table 1.This gives a more convincing picture than reporting sparsity alone. In particular, the results on page 8 suggest that prox-mix reaches very strong compression regimes while keeping Dice close to the original nnU-Net.
The 3D results are particularly important because that is where nnU-Net’s computational burden is often most restrictive.
- 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.
My main concern is novelty relative to existing pruning literature. Sparse-group regularization, proximal optimization, and structured pruning are all established ideas, and the manuscript itself cites prior work on proximal training and structured sparsity. The main novelty here seems to be the adaptation of these ideas to nnU-Net and the framing as “self-configuration from an optimization angle.” That is interesting, but the conceptual advance over existing structured pruning/proximal methods feels somewhat incremental.
- 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
This is a solid and well-presented paper with a nice optimization perspective on nnU-Net compression. I think it would become stronger with three additions: (1) include a reduced-width / reduced-capacity nnU-Net baseline trained from scratch at similar compute; (2) expand the experimental validation to more datasets or at least provide stronger discussion of generality; and (3) clarify hyperparameter and threshold selection so the reported Pareto frontier is easier to interpret from a model-selection standpoint.
- 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?
I am slightly positive on this paper. The problem is relevant, the method is principled, and the results are strong enough to be interesting. In particular, the combination of very high structural compression and near-baseline Dice on both ACDC and BraTS is compelling, and the inclusion of real runtime measurements is a plus. My reservation is that the paper does not yet fully prove that this is the best way to obtain efficient nnU-Nets, rather than one reasonable pruning-based route. The novelty over prior structured pruning and proximal sparsity work is moderate, and the missing smaller-model baseline leaves an important question unanswered. That keeps me at weak accept rather than a stronger score.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 #3
- Please describe the contribution of the paper
This work formulates nnU Net training as a proximal optimization problem with mixed ℓ₁ and group ℓ₂ regularization to encourage structural sparsity. The approach removes entire channels during training, unifying weight and structure learning into a single process, resulting in slim nnU Net architectures.
- 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 manuscript is well-structured.
- The motivation is clearly explained.
- 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.Pruning is a well-established technique for network compression. Both single-granularity (e.g., unstructured, semi-structured, structured) and mixed-granularity pruning have been extensively studied. (1) The related work would benefit from including a discussion of other mixed granularity pruning techniques. (2) The reviewer is not fully convinced why the proposed method is specifically investigated on nnU Net, as both the addressed problem and the proposed approach appear general and applicable to other networks and tasks as well. (3) A comparison with state of the art mixed granularity pruning methods should be provided in Section 3. 2.The use of L1/L2 regularization in the loss function to encourage unstructured or structured sparsity, as well as channel level coarse grained pruning, has already been explored in prior works. Consequently, the novelty of the proposed method appears somewhat limited.
3.The experiments are limited to nnU‑Net. It is recommended to evaluate it on more SOTA networks. In addition, the method should be compared with other leading pruning approaches, including mixed‑granularity methods.
4.More implementation details such as training hyperparameters and pre-processing strategies should be provided to ensure reproducibility.
- 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.
(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 major concerns include 1) the relatively limited novelty of the proposed approach, 2) the narrow experimental scope, restricted to nnU-Net without comparison to SOTA pruning methods, and 3) the insufficient implementation details.
- 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.
Thank you for the response. After reading the rebuttal and other reviews, I would like to keep my current rate.
Author Feedback
We thank all reviewers for their constructive feedback. We are encouraged by the positive remarks that the optimization perspective is novel (R1) and clean (R2), nnU-Net reconfiguration is high relevant and practical (R2), and that our results are strong and comprehensive (AC) by including FLOP and runtime gains. We address the main concerns below. Novelty of the work: We appreciate R1’s comment that reformulation of nnU-Net configuration from an optimization angle is a major novelty. nnU-Net is the cornerstone of medical image segmentation, but its computational cost is high. Optimization of nnU-Net configuration is meaningful from both theoretical and practical angles, but not well studied. In standard nnU-Net, self-configuration is rule-based, while existing pruning works were not compatible with nnU-Net: they typically adopt training-pruning-retraining pipelines, which are heuristic-driven, iterative, and sub-optimal from the optimization angle. In contrast, our work is built on the proximal optimization theories, and the pruning is done in one training trajectory (Fig. 1). It is theoretically grounded, and practically keeps the nnU-Net-v2 framework intact: no dynamic topology changes, pruning-retraining cycles, or ad hoc feature-analysis are needed; only a proximal step is inserted after the standard gradient update. This offers both theoretical and practical advantages over existing work.
Smaller-model baseline: We thank R2 for this important remark. We address it from two angles: first, without the proposed proximal optimization, we could not reach the specific configuration; the work would then change to expensive neural architecture search (NAS); second, over-parametrization is crucial for DL training to find good solutions, and prior work has shown that starting from small networks cannot match the results of starting from large networks and then pruning. We have done additional experiments that confirmed this: a randomly initialized small nnU-Net with matched FLOPs showed Dice drops of 1.91% on ACDC and 1.68% on BraTS. The difference is significant in medical image segmentation benchmarks. We will add this discussion to revision.
Reproducibility and implementation details: We will release the code upon acceptance.
Hyperparameter and pruning-threshold selection (Meta, R2, R3): The sparsity weights λg and λ1 can be selected on the validation set according to the desired Dice-FLOP trade-off. In our experiments, favorable settings were consistent: both ACDC and BraTS worked well around λg = 1e-3, with λ1 in the range of 5e-4 to 1e-3.The pruning threshold τ is selected after training as an operating point. Sweeping τ gives the validation Dice-FLOP Pareto curve, from which one can choose a compact model according to the acceptable Dice drop. The final operating point can be chosen from the validation Dice-FLOP Pareto curve to achieve strong FLOP reduction with minimal Dice drop; in our experiments, this typically corresponded to τ in the range of 0.05–0.1 relative to the layer-wise maximum group norm. We will add the details to the revision.
Focus on nnU-Net and benchmarking (R3): The paper indeed focuses on nnU-Net self-configuration, rather than generic pruning of architecture. This is a deliberate choice: despite rapid developments, nnU-Net remains the de facto framework for medical segmentation. Therefore, making nnU-Net more efficient, while keeping the original implementation intact, has broad practical impact, especially for 3D cases. For benchmarking, we included state-of-the-art baselines for U-Net and ablation studies: LTH, STAMP, Sauron, GD-mix, and prox-group within nnU-Net-v2 framework. It covered unstructured, structured, feature-based pruning, and variants of proximal sparsity. We showed that prox-mix achieved the best accuracy-efficiency tradeoff (Fig.2), providing a simple, principled, and nnU-Net-compatible way for boosting efficiency without losing performance.
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 makes nnUNet more efficient and the empirical. The results are strong, especially the large structured compression, FLOP reduction, and actual runtime gains while keeping Dice close to the original nnUNet. However there still exist some concern: sparse group regularization, proximal optimization, and structured pruning are established ideas, and the paper’s contribution is mainly adapting and framing them for nnUNet. For rebuttal, the authors should clarify the novelty over prior structured/proximal pruning, justify the contribution, explain hyperparameter and pruning-threshold selection.
- 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 proposes Slim nnU-Net, which formulates nnU-Net compression as a proximal optimization problem with mixed element-wise and group-wise sparsity regularization. The problem is practically relevant, since nnU-Net remains widely used but can be computationally expensive, especially in 3D segmentation.
The main strength of the submission is its empirical evaluation. The method is tested on both 2D cardiac MRI segmentation and 3D brain tumor segmentation, and the submitted manuscript reports strong accuracy-efficiency trade-offs, with substantial reductions in model size and FLOPs while maintaining near-baseline Dice performance. The reported CPU and GPU inference measurements further support the practical value of structured compression.
I agree with the reviewers that the methodological novelty is moderate, since proximal optimization and structured sparsity are established techniques, and broader comparison with alternative compact nnU-Net designs would strengthen the work. Nevertheless, the adaptation to nnU-Net is well motivated, technically sound, and experimentally convincing. Overall, I consider the paper above the acceptance threshold.
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.
As pointed out by the reviewers, combining nnU-Net with network pruning is an interesting idea with practical importance. Although I agree with Reviewer #3 that the theoretical contribution of the pruning component is relatively weak, the integration of nnU-Net with pruning appears to be novel, and its practical advantages could be substantial. Considering the rebuttal as well, I recommend accepting this paper.
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.
The paper received mixed reviews, with both positive and negative opinions. After carefully reading the rebuttal and the reviewers’ comments, I believe that the authors have addressed most of the concerns raised during the review process. Although some issues remain not fully resolved, I find the work interesting and the remaining weaknesses do not constitute sufficient grounds for rejection. Therefore, I recommend accepting the paper, and I encourage the authors to further improve the manuscript in the final version.