List of Papers Browse by Subject Areas Author List
Abstract
While U-Net architectures remain the gold standard for medical image segmentation, their deployment in resource-constrained environments demands aggressive model compression. However, finding an optimally efficient configuration is computationally prohibitive, typically requiring exhaustive train-and-evaluate cycles to find the smallest model that maintains peak performance. In this paper, we introduce a training-free selection framework to automatically identify ultralightweight, dataset-specific U-Net configurations directly at initialization. We observe that systematically scaling down U-Net channel width induces a sharp transition from a stable performance plateau to representational capacity collapse. To pinpoint this boundary without training, we propose a Jacobian-based sensitivity metric that scores discrete, width-capped U-Net variants using a small set of unlabeled images. By analyzing the total variation of this sensitivity curve, we isolate the smallest stable configuration, which we denote as XTinyU-Net. Evaluated across six diverse medical datasets within the nnU-Net framework, XTinyU-Net achieves segmentation accuracy comparable to the heavy nnU-Net baseline with 400x-1600x fewer parameters, and outperforms contemporary lightweight architectures while utilizing 5x-72x fewer parameters. Code is publicly accessible at https://github.com/alvinkimbowa/nntinyunet.git.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6032_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/alvinkimbowa/nntinyunet.git
Link to the Dataset(s)
N/A
BibTex
@InProceedings{KimAlv_XTinyUNet_MICCAI2026,
author = { Kimbowa, Alvin AND Heidari, Moein AND Liu, David AND Hacihaliloglu, Ilker},
title = { { XTinyU-Net: Training-Free U-Net Scaling via Initialization-Time Sensitivity } },
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 novel method for training-free selection of an optimally sized UNet configuration at initialization. The motivation behind this is that, usually, with UNet configurations there is a long plateau of performance in comparison to model size, followed by a sudden drop in performance. By doing so, the authors manage to quickly find the best UNet configurations that have significantly fewer parameters while still achieving similar performance, and also outperform particularly minimally designed UNet 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.
- Interesting idea that has potential for real-world acceleration in finding the best UNet configuration for a given task.
- Well-written paper. Easy to follow, and there was clearly effort put into communicating the intuition.
- Large number of datasets and baselines.
- Also a large number of evaluated configurations to ensure the proposed method works.
- Overall, enough evidence to convince me that this method generally works.
- 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.
- It is not fully clear whether UNeXt, Tiny U-Net, etc. , have been evaluated within the nnUNet framework. If not, this is not a fair comparison, and it is then no surprise to see these performance differences.
- Considering the qualitative results, all of these methods perform well enough; the main question of interest here would have been robustness. If we reduce parameters significantly but then lose robustness, we do not really gain anything.
- Regarding Figure 4b (Effect of initialization), the authors mention that it is quite robust against initialization, but I would not agree with this. Jumping between three model configurations that are orders of magnitude apart is definitely not robust. What if you have a configuration where the peak (assuming there always is one) is in between two points?
- 4b) This would have benefited from a more extensive statistical analysis. Right now, it is not clear if these few data points give us the right idea in terms of robustness.
- 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 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
The research question is repeated too often.
- 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?
Overall, a well-written paper with a method that is mostly convincing in terms of its effectiveness. It lacks some statistical analysis on the initialization to fully demonstrate that it is broadly applicable; also, some details regarding baselines remain unclear.
- 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’s main contribution is to provide a training-free method for selecting the smallest UNet (from a parameter-count perspective) without affecting the model’s absolute performance. The XTinyUNet performs similarly to a standard nnUNet with orders of magnitude fewer parameters and GFLOPS.
- 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 authors propose a training-free method for optimizing UNet architecture by scaling its width (channel count). By utilizing a Jacobian sensitivity measure, the approach identifies the ‘empirical collapse boundary’, the point of minimum parameter count just before model accuracy sharply declines. This allows for the pre-selection of the most efficient model size without compromising performance metrics such as Dice or HD95
- 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 paper is technically solid and, within the limited space, has provided sufficient empirical evidence to support its methodology. Additional comments are provided in section 12.
- 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 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
- On experimental setup:
- While I acknowledge that this paper is focused on segmentation, and segmentation datasets are limited in size, do the authors foresee that this approach can be extended to other encoder-based architectures for classification for example? If so, then the sensitivity measure would need to be modified (see questions below for clarification on the RMS estimate). ○ If dataset size increases substantially, how would this impact the selection process? Would this approach still hold? I am assuming most of the datasets are of the size on the order of 10^3 , if it’s scaled to 10^6 would this approach still hold?
- On evaluation:
- The paper compares absolute performance as measured by Dice / HD95, what about calibration? NLL, Brier Score?
- Questions:
- Section 2.2 for input-output sensitivity. The authors state that a gradient of the model’s output is taken with respect to its input at “random initialization” using a RMS estimate: ○ the RMS estimate is based on the root mean square difference of the output and the input, from a randomly initialized network? ○ So does that mean all the dataset inputs (images) are fed forward through the network to compute this sensitivity? And then the sensitivity is estimated based on the entire dataset? As in a whole epoch? ○ And I presume that the weights are not updated (since it is training-free), it’s a simple feedforward step and gradient computation with respect to RMS(input, output)?
- If XTinyUNet is defined with the k* at the maximum collapse boundary, which is at the transition from stable plateau to a degradation in performance with smaller models (post collapse) does that mean the model has a finite capacity limited by the given dataset? Is there a relationship between the relative size of the model and relative data complexity (size) that can be deduced from this collapse boundary definition?
- What is the rationale behind this optimum point of performance relating to the size of the dataset?
- How does this initialization method relate to training a sparse network which can significantly reduce number of parameters, albeit can have a higher flops? Ref: Rigging the Lottery: Making All Tickets Winners (Evci, et al 2019, ICML).
- 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?
Solid paper, well written. Clear motivation, methodology, and claims are well-supported with extensive empirical evidence and solid experimental design.
- 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
The paper presents XTinyU-Net, a training-free framework for automatically selecting ultra-lightweight U-Net configurations for medical image segmentation through initialization-time sensitivity analysis. The central contribution of the work is not the proposal of a new segmentation architecture in the conventional sense, but rather a new methodology for model scaling and capacity selection. Instead of designing yet another compact encoder-decoder network or relying on expensive train-and-evaluate sweeps across many candidate models, the paper asks a practical and important question: how small can a standard U-Net be made for a specific dataset before segmentation performance collapses. The authors propose that this boundary can be estimated directly at random initialization without full training, thereby converting lightweight architecture design into a training-free model selection problem. The first major contribution is the empirical identification of a consistent two-regime behavior when systematically shrinking U-Net width. By progressively reducing the maximum channel capacity of nnU-Net style U-Net models, the authors observe that segmentation accuracy across multiple datasets remains relatively stable over a broad range of compressed configurations, forming a performance plateau. However, beyond a dataset-dependent threshold, performance drops sharply once representational capacity becomes insufficient. The paper refers to this transition as the boundary between pre-collapse and post-collapse regimes. This is an insightful observation because it suggests that many U-Net models are substantially over-parameterized for specific medical datasets, and that aggressive compression is possible until a critical tipping point is reached. Rather than treating model scaling as a smooth trade-off, the paper frames it as a phase-transition style phenomenon. The second and most important methodological contribution is a Jacobian-based sensitivity metric for detecting this collapse boundary at initialization time. For each candidate U-Net variant, the model is randomly initialized and a small batch of unlabeled input images is passed through the network. The gradient of the summed outputs with respect to the input image is computed, and the norm of this Jacobian is used as a scalar sensitivity measure. The intuition is that under-parameterized models near capacity collapse become increasingly unstable and more sensitive to small input perturbations. By measuring how this sensitivity changes across progressively smaller U-Net variants, the authors derive a discrete sensitivity curve over the model family. This is a creative use of initialization-time network dynamics for practical segmentation model selection. The third contribution is the proposed collapse-boundary detection rule based on total variation of the sensitivity curve. Rather than simply ranking architectures by the raw Jacobian score, the method analyzes the change in sensitivity between neighboring model scales. It searches for the split point where the large-model region remains stable while the small-model region exhibits rapidly increasing variation. The selected architecture is the smallest model that still lies on the stable side of this transition. This chosen model is termed XTinyU-Net. Conceptually, this is significant because the method does not seek the globally highest proxy score, but instead estimates the onset of representational failure and selects the last safe compression point. A further contribution is that the proposed framework is fully training-free during model selection. Traditional lightweight architecture optimization requires training many candidate models, often consuming large computational resources and making deployment adaptation slow or impractical. By contrast, XTinyU-Net only requires forward and backward passes at initialization using a small number of unlabeled samples. This dramatically reduces the cost of selecting a dataset-specific compact architecture. The approach is therefore highly practical for environments where compute, time, or annotation budgets are limited. Another important contribution is the demonstration that standard U-Net families, when properly scaled, can outperform many purpose-built lightweight segmentation networks. The paper compares XTinyU-Net against UNeXt, CMUNeXt-S, TinyU-Net, and the original nnU-Net across six public medical imaging datasets spanning multiple modalities and tasks: BUS-BRA breast ultrasound, EchoNet-Dynamic echocardiography, ISIC 2018 dermoscopy, FiVES retinal vessels, BraTS brain tumor MRI, and ACDC cardiac MRI. This broad evaluation is a strength because it shows that the proposed framework is not tied to one modality or anatomy. Quantitatively, XTinyU-Net achieves performance comparable to the full nnU-Net baseline while using dramatically fewer parameters and FLOPs. Depending on the dataset, the reported reductions range from approximately 400 times to 1600 times fewer parameters. For example, on BUS-BRA, XTinyU-Net reportedly attains 90.64 Dice compared with 89.90 for nnU-Net while reducing parameters from over 33 million to roughly 20 thousand. Similar strong efficiency-performance trade-offs are shown across EchoNet-Dynamic, FiVES, BraTS, and ACDC. These results support the paper’s core thesis that dataset-specific width scaling of a canonical architecture can be more effective than designing entirely new lightweight networks. The paper also contributes a comparison with existing zero-cost NAS metrics such as NASWOT and SWAP. The authors argue that standard zero-cost metrics tend to scale monotonically with model size and therefore fail to identify the abrupt transition boundary relevant for intra-family U-Net scaling. In contrast, the proposed Jacobian sensitivity metric exhibits a distinct peak near the collapse threshold. This comparison is useful because it clarifies that the problem here is not generic architecture ranking, but regime-change detection within a structured model family. At a broader level, the conceptual contribution of the paper is the reframing of lightweight segmentation model design. Instead of asking which custom architecture should be invented next, the paper asks whether a well-established architecture such as U-Net can simply be scaled intelligently using initialization-time signals. This is an appealing perspective because it emphasizes principled resource adaptation over continual architectural proliferation. Overall, the main contribution of the paper is a practical and novel training-free methodology for selecting the smallest stable U-Net for a given medical segmentation dataset. Through the discovery of a performance plateau followed by abrupt collapse, the introduction of Jacobian sensitivity based collapse detection, and extensive validation across six datasets, the paper provides an efficient alternative to manual model search and suggests that canonical U-Net architectures remain highly competitive when scaled appropriately.
- 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 paper has several notable strengths, even though I remain below the acceptance threshold overall. The strongest aspect of the submission is that it focuses on a highly practical and often overlooked problem in medical image segmentation: efficient model configuration selection. A large portion of recent literature concentrates on proposing new architectures, yet in real deployment settings many practitioners still rely on strong baselines such as U-Net or nnU-Net and need principled ways to reduce model size for resource-constrained environments. By addressing the question of how aggressively a standard U-Net can be compressed for a specific dataset before performance degrades, the paper tackles a relevant and deployment-oriented problem rather than introducing architecture novelty for its own sake. A major strength is the conceptual reframing of lightweight model design. Instead of asking which new compact network should be invented, the paper proposes that a canonical architecture may already be sufficient if scaled correctly. This is an interesting perspective because it challenges the common assumption that efficiency must come from handcrafted lightweight blocks, token mixers, or architectural redesign. The suggestion that intelligent configuration search can outperform bespoke compact models is both practical and thought-provoking for the community. Another important strength is the proposed training-free model selection strategy. The method uses initialization-time Jacobian sensitivity to estimate the smallest stable U-Net configuration without fully training every candidate model. This is appealing because exhaustive train-and-evaluate sweeps are computationally expensive, especially in medical imaging where experiments often involve large volumes, heavy augmentations, and cross-validation protocols. A selection mechanism that only requires random initialization plus a few forward and backward passes can significantly reduce experimentation cost. Even if the method requires further validation, the overall idea is novel and operationally useful. The empirical observation of a two-regime scaling pattern is also a meaningful contribution. The paper shows that as U-Net width decreases, segmentation accuracy often remains on a broad plateau across many compressed configurations, followed by a sharp collapse once capacity becomes too small. This phenomenon is interesting because it suggests that many segmentation models are over-parameterized relative to dataset complexity. Framing model scaling as a plateau-to-collapse transition rather than a smooth monotonic degradation is a useful insight that could motivate future work in compression and neural scaling laws for medical imaging. A further strength is the use of Jacobian sensitivity as a collapse detector rather than simply as a generic zero-cost ranking metric. Existing zero-cost NAS approaches are often designed to compare unrelated architectures, whereas this paper focuses on a structured family of width-scaled U-Nets. The idea of using local input-output sensitivity to identify instability near the representational boundary is creative. Even if the theoretical underpinnings could be strengthened, it is an original application of initialization-time analysis to dense prediction model scaling. The breadth of evaluation is another positive aspect. The paper reports experiments across six public datasets spanning diverse imaging modalities and tasks: breast ultrasound, echocardiography, dermoscopy, retinal vessel segmentation, brain tumor MRI, and cardiac MRI. This is stronger than many papers that validate only on one or two benchmarks. The diversity of modalities suggests the method may capture a general property of U-Net capacity scaling rather than overfitting to one domain. The reported efficiency gains are also impressive. XTinyU-Net is claimed to achieve performance comparable to nnU-Net while using approximately 400x to 1600x fewer parameters depending on the dataset, along with substantial FLOP reductions. If these results hold under broader testing, they would be practically meaningful for edge deployment, mobile inference, portable imaging systems, and institutions with limited hardware resources. The ability to reduce a standard nnU-Net style model from tens of millions of parameters to tens of thousands while maintaining competitive Dice scores is an attractive result. Another strength is the comparison against contemporary lightweight baselines such as UNeXt, CMUNeXt-S, and TinyU-Net. The paper does not only compare to the heavy nnU-Net upper bound but also to models specifically designed for efficiency. This makes the evaluation more relevant because it tests whether architecture search through scaling can compete with dedicated lightweight design. The ablation analysis adds additional value. The authors test robustness of the proposed selection criterion to batch size and random initialization, and compare against existing zero-cost metrics such as NASWOT and SWAP. Showing that the chosen architecture remains relatively stable under different initialization seeds and even with a small number of input samples is encouraging. This kind of robustness analysis is important for methods that rely on initialization statistics. The paper is also clearly written and easy to follow. The motivation is intuitive, the method is mathematically simple, and the main figures communicate the plateau-collapse phenomenon effectively. The overall narrative is coherent: identify redundancy, detect collapse boundary, and choose the smallest safe model. From a translational standpoint, the work has potential clinical relevance because resource-efficient segmentation models are increasingly important. Portable ultrasound devices, point-of-care scanners, low-power hospital systems, and real-time embedded applications often cannot support heavy nnU-Net style deployments. A training-free route to generate compact dataset-specific models could therefore be useful in practice.
Finally, the paper has strong practical elegance. It does not rely on complex multi-stage distillation, pruning, quantization, or expensive NAS pipelines. Instead, it proposes a simple structured family of U-Net variants and a lightweight sensitivity criterion. This simplicity may help adoption and reproducibility if validated further. Overall, although I have concerns that place the paper below threshold, the submission has genuine strengths: it targets a practical problem, offers a creative training-free selection framework, provides broad empirical evaluation, and challenges the assumption that new lightweight architectures are always necessary. These aspects make it an interesting and worthwhile contribution.
- 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 paper has several strengths, but there are also important weaknesses that prevent me from rating it above the acceptance threshold in its current form. My main concerns relate to novelty, rigor of validation, fairness of comparisons, methodological assumptions, and practical deployment claims. The first major weakness is limited methodological novelty. The paper does not introduce a fundamentally new segmentation architecture, learning objective, or optimization framework. Instead, it proposes selecting channel-scaled U-Net variants using an initialization-time proxy score. While practical, this places the contribution closer to model selection heuristics or zero-cost NAS adaptation than a substantial new medical image segmentation method. Training-free architecture scoring at initialization has already been explored extensively in the broader NAS literature through methods such as NASWOT, SynFlow, SNIP, Zen-NAS, and SWAP. The paper acknowledges some of these works, but the methodological jump beyond prior zero-cost proxies is moderate. The main change is applying Jacobian sensitivity to detect a collapse point within a family of width-scaled U-Nets rather than ranking unrelated architectures. This is useful, but somewhat incremental rather than a major conceptual advance. A second weakness is that the central claim of a plateau-then-collapse scaling law is not fully established. The paper shows six datasets where Dice remains stable over a range of widths and then drops sharply. However, six datasets are still limited relative to the diversity of medical segmentation tasks. It remains unclear whether this behavior consistently holds for highly imbalanced segmentation, large 3D volumetric tasks, noisy labels, pathology images, longitudinal imaging, or highly multi-class dense labeling. The claim may be valid for the tested settings, but broader generalization is not yet sufficiently demonstrated. Relatedly, the datasets are somewhat heterogeneous but mostly moderate-scale and standard benchmarks. The study lacks validation on larger and more demanding tasks where nnU-Net is widely used, such as MSD benchmarks, full 3D abdominal organ segmentation, large CT cohorts, or multi-center clinical data. Since the method is framed around nnU-Net scaling, stronger evidence on more canonical nnU-Net challenge datasets would strengthen the case. A third weakness is that the proposed Jacobian sensitivity metric lacks strong theoretical grounding. The paper argues that under-parameterized networks become more sensitive to input perturbations near the capacity boundary, and therefore Jacobian norms can detect collapse. While intuitive, this connection is not rigorously justified. Jacobian norms at random initialization are known to depend on initialization scheme, activation function, normalization layers, depth, input resolution, and scaling constants. It is unclear whether the measured peaks truly correspond to representational insufficiency or simply reflect optimization dynamics specific to the chosen implementation. A deeper theoretical or empirical analysis would help justify why this metric should generalize. Another concern is sensitivity to implementation details. Since the method depends on random initialization behavior, seemingly minor changes such as He versus Xavier initialization, normalization placement, residual versus plain blocks, deep supervision, or preprocessing may alter the sensitivity curve. The paper shows some robustness to seeds and batch size on one dataset, but does not study robustness across initialization schemes or architecture variants. This limits confidence in broader applicability. The fairness of comparisons also deserves scrutiny. XTinyU-Net is compared against UNeXt, CMUNeXt-S, TinyU-Net, and nnU-Net, but these baselines may not all be equally optimized under identical training budgets. Some lightweight methods may require architecture-specific tuning rather than default nnU-Net style settings. If all baselines are trained under one common recipe that favors standard U-Net style models, then the comparison may advantage XTinyU-Net. Stronger fairness would require reproducing each baseline under its recommended protocol or reporting official numbers alongside standardized retraining. A related issue is that only a limited set of lightweight baselines is included. There are many recent efficient medical segmentation models such as Mobile U-Net variants, EfficientUNet derivatives, lightweight transformers, pruning or distillation approaches, MedNCA, LeanUNet, and other task-specific compact architectures. The comparison space therefore feels selective rather than exhaustive. This weakens the claim that XTinyU-Net consistently outperforms the state of the art in lightweight segmentation. Another weakness is the use of parameter count as a dominant efficiency metric. While reducing parameters from 33M to 20K is striking, parameters alone do not fully capture deployment efficiency. Real inference speed depends on memory access patterns, hardware utilization, operator fusion, precision mode, and tensor shapes. Very narrow networks can sometimes underutilize GPUs or NPUs. Although FLOPs are reported, wall-clock latency, throughput, CPU or mobile inference time, and memory footprint are not comprehensively benchmarked. Since the paper is motivated by deployment, practical runtime evidence would be important. The clinical feasibility narrative is also somewhat indirect. The paper argues that lightweight segmentation is useful for portable and decentralized settings, which is true in principle, but no actual embedded deployment, point-of-care experiment, mobile benchmark, or clinician-facing use case is demonstrated. Thus, the practical significance is plausible but not directly validated. Another concern is that the method optimizes only width scaling while keeping depth and topology fixed. This simplifies the search space, but may also limit achievable efficiency. Many successful lightweight models derive gains from changing kernel sizes, stage depth, skip pathways, separable convolutions, decoder simplification, or attention placement. Restricting the search to width alone may miss better Pareto-optimal architectures. Therefore, the conclusion that a properly scaled U-Net is superior to custom lightweight designs may be overstated. The reported results also raise some questions. In several datasets XTinyU-Net slightly exceeds nnU-Net despite having orders of magnitude fewer parameters. While possible due to reduced overfitting, such dramatic compression with equal or better performance across many datasets is surprising and merits deeper analysis. Confidence intervals are given, but only three runs are used. More seeds, cross-validation, or statistical significance testing would strengthen confidence that these gains are robust rather than variance-driven. Reproducibility is moderate but incomplete. The method description is understandable, yet exact implementation details for candidate family generation, sensitivity computation overhead, image sampling strategy, preprocessing consistency, and integration into nnU-Net versions could be more explicit. Since the paper claims training-free rapid selection, reporting actual selection runtime compared with conventional sweeps would also be useful. Finally, the paper’s impact may be narrower than it first appears. If the method mainly works for U-Net width scaling inside nnU-Net-like pipelines, then it is a useful engineering tool but not necessarily a broad methodological advance for MICCAI. Overall, the paper is interesting and practically motivated, but the novelty is moderate, the evidence base is narrower than the claims suggest, comparisons could be stronger, and the theoretical justification remains underdeveloped. These issues collectively place the work below threshold for me in its current form.
- 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
The paper addresses a practical and relevant problem in medical image segmentation: how to obtain highly efficient models without repeatedly training large numbers of candidate architectures. This is an important topic for the MICCAI community, particularly as deployment increasingly moves toward portable devices, edge systems, low-resource clinics, and real-time applications where full nnU-Net style models may be impractical. The manuscript is clearly written, easy to follow, and presents a coherent narrative from motivation through experiments. One of the most interesting aspects of the paper is the decision to challenge the current trend of continual lightweight architecture redesign. Instead of proposing another custom encoder-decoder, the paper asks whether a standard U-Net family can simply be scaled intelligently. This is a refreshing and practically useful perspective. In many real-world settings, practitioners prefer trusted and interpretable baselines rather than highly specialized architectures. If a standard U-Net can be reduced dramatically while preserving performance, this could be valuable. The observation of a broad performance plateau followed by sharp collapse under width scaling is also thought-provoking. If robust, this suggests that many segmentation models are substantially over-parameterized for certain datasets. The idea of selecting the last stable model before collapse is intuitive and may inspire future work on scaling laws, dataset complexity estimation, and training-free model compression for medical imaging. To further strengthen the paper, I encourage the authors to more clearly position the contribution relative to prior zero-cost NAS and initialization-based scoring literature. Methods such as NASWOT, SynFlow, SNIP, Zen-NAS, and SWAP have already explored training-free architecture assessment. The key novelty here seems to be collapse-boundary detection within a structured family of width-scaled U-Nets rather than generic architecture ranking. Making this distinction explicit and emphasizing why intra-family collapse detection is different from standard NAS ranking would improve clarity and help readers better appreciate the contribution. The empirical section would also benefit from stronger validation on larger and more canonical nnU-Net use cases. Since nnU-Net is widely associated with volumetric CT and MRI benchmarks, adding experiments on representative 3D datasets such as MSD tasks or larger multi-organ segmentation problems would considerably strengthen the paper. Currently, much of the evidence comes from 2D or moderate-scale benchmarks. Demonstrating that the same plateau-collapse behavior and selection rule extends to full 3D pipelines would increase confidence in general applicability. I would also encourage deeper analysis of the proposed Jacobian sensitivity signal. The current intuition is reasonable, but stronger evidence would help. For example, testing sensitivity under different initialization schemes, normalization choices, activation functions, or optimizer settings could reveal whether the selected boundary is robust or tightly coupled to a specific implementation. If the signal is stable across such changes, that would significantly strengthen the contribution. Additional runtime analysis would also be valuable. Since the paper is motivated by efficiency, readers would benefit from wall-clock inference latency, memory footprint, CPU benchmarks, mobile deployment tests, or energy usage where possible. Parameter count and FLOPs are useful but incomplete proxies for deployment efficiency. The comparison set could be expanded. While UNeXt, CMUNeXt-S, TinyU-Net, and nnU-Net are reasonable references, the field now includes several recent compact segmentation models such as LeanUNet, MedNCA, pruning-based methods, and efficient transformer hybrids. Including a broader comparison would make the claims of superiority more convincing. Some of the quantitative results are impressive enough that they deserve deeper interpretation. In several datasets, XTinyU-Net approaches or slightly exceeds nnU-Net while using dramatically fewer parameters. This is possible, particularly if the baseline is over-parameterized, but additional discussion would help readers understand why. For instance, are gains due to better regularization, reduced overfitting, dataset simplicity, or redundancy in standard width schedules? Such analysis would improve scientific insight beyond benchmark numbers. From a reproducibility perspective, the paper would benefit from more implementation details in the final version. Please consider including exact candidate width schedules, preprocessing steps, random seed policy, sensitivity computation time, hardware used for scoring, and integration specifics with nnU-Net versions. Since the method is intended to save practitioner time, reporting the total search time versus conventional train-and-sweep pipelines would be especially compelling.
The broader impact of the work may be strongest if framed as a practical systems contribution rather than a universal replacement for architecture innovation. The message that “well-scaled standard models can rival bespoke lightweight designs” is valuable, but it may be more nuanced than absolute superiority. Presenting the method as a strong baseline and efficient deployment tool could resonate well with readers. Overall, this is a thoughtful and practically motivated paper with a clean idea and encouraging results. The strongest contributions are the reframing of lightweight model design as training-free capacity selection, the empirical identification of collapse behavior under width scaling, and the possibility of obtaining strong segmentation performance with extremely compact standard architectures. With broader validation, stronger theoretical grounding, expanded comparisons, and more deployment-oriented evidence, the paper could make a meaningful contribution to efficient medical AI systems. I encourage the authors to continue developing this direction.
- 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?
I assigned a reject because, although the paper is practical, clearly written, and presents an interesting efficiency-oriented idea, I do not believe the current submission reaches the level of methodological novelty, empirical rigor, and broad impact typically expected for acceptance at MICCAI. The paper is not without merit, but in its present form I found the overall contribution below threshold.
One positive aspect is that the paper addresses a real and important problem. Many medical image segmentation systems rely on large U-Net or nnU-Net style models that can be difficult to deploy in low-resource environments, portable devices, or latency-sensitive settings. Reducing the computational burden of segmentation models is therefore highly relevant. I also appreciate the paper’s practical framing: instead of proposing another handcrafted lightweight architecture, it asks whether a standard U-Net can simply be scaled intelligently. This is a useful perspective and likely of interest to practitioners.
However, my primary concern is limited methodological novelty. The core contribution is not a new segmentation model or learning framework, but a training-free selection rule for width-scaled U-Net variants using an initialization-time Jacobian sensitivity signal. While useful, this feels closer to an engineering heuristic or adaptation of zero-cost neural architecture search ideas than a substantial new methodological advance in medical image computing. Prior literature has already explored training-free architecture scoring and initialization-based proxies through methods such as NASWOT, SynFlow, SNIP, Zen-NAS, and SWAP-NAS. The present work applies a related philosophy to a structured family of U-Net models and uses collapse-point detection rather than generic ranking. This is a reasonable extension, but for a selective venue, I found the novelty moderate rather than strong.
A second major factor was insufficient evidence for generality of the central claim. The paper argues that U-Net width scaling exhibits a stable performance plateau followed by abrupt collapse, and that this transition can be detected without training. While the experiments are interesting, the evidence is based on six datasets, many of which are relatively standard and moderate-scale. It remains unclear whether the same behavior would hold consistently for more complex settings such as full 3D volumetric segmentation, highly multi-class organ delineation, pathology gigapixel imagery, severe class imbalance, multi-center clinical cohorts, or noisier annotations. Since the paper makes broad claims about dataset-adaptive scaling, stronger validation across more challenging scenarios would be needed.
Relatedly, because the method is positioned within the nnU-Net ecosystem, I expected stronger experiments on representative large-scale nnU-Net style benchmarks, such as MSD tasks or demanding CT/MRI segmentation problems. Demonstrating the same selection principle on these canonical use cases would significantly strengthen confidence and relevance to the MICCAI audience.
Another concern is the theoretical grounding of the proposed Jacobian sensitivity proxy. The intuition that under-parameterized networks become unstable near the collapse boundary is plausible, but the paper does not convincingly establish why the chosen metric should robustly predict segmentation capacity across architectures and datasets. Jacobian norms at initialization can depend strongly on initialization scheme, normalization layers, activation functions, depth, resolution, and implementation details. The paper includes some robustness analysis, but I did not find sufficient evidence that the signal is fundamentally linked to segmentation representational limits rather than to specific design choices.
The comparison space also influenced my score. While the paper compares against UNeXt, CMUNeXt-S, TinyU-Net, and nnU-Net, the set of baselines feels limited relative to the broader landscape of efficient segmentation models. There are numerous recent compact CNNs, lightweight transformers, pruning methods, distillation methods, and optimized U-Net variants. A broader and stronger comparison would be necessary before concluding that properly scaled vanilla U-Nets outperform specialized lightweight architectures.
I was also not fully convinced by the efficiency evidence. The paper emphasizes dramatic reductions in parameter count and FLOPs, which are valuable metrics, but practical deployment efficiency usually requires wall-clock latency, throughput, memory footprint, CPU performance, mobile hardware benchmarking, or energy measurements. Extremely small models do not always translate linearly into real deployment speedups. Since the work is motivated by resource-constrained deployment, more direct systems evidence would have strengthened the practical case.
Some reported results are impressive enough that they also warrant deeper analysis. In several datasets, XTinyU-Net matches or slightly exceeds nnU-Net despite having orders of magnitude fewer parameters. This is possible if the baseline is substantially over-parameterized, but repeated confirmation across many datasets suggests the need for deeper explanation, stronger statistical testing, and more extensive runs. Only three runs were reported in many settings, which may be insufficient to fully assess variance.
Finally, while the paper is polished and practically motivated, the broader impact appears somewhat narrower than the title may suggest. If the method mainly works as a width-selection tool for U-Net style models under nnU-Net-like training pipelines, then it is a useful engineering contribution, but perhaps not a sufficiently broad algorithmic advance for acceptance in a highly competitive conference.
Overall, I view the paper as thoughtful and potentially useful, but currently below threshold due to moderate novelty, limited validation breadth, incomplete efficiency evidence, and insufficient theoretical justification of the proposed proxy. My reject score reflects that the paper has promise, but I was not yet convinced that it makes a strong enough contribution relative to the expected standard for MICCAI acceptance.
- 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
Author Feedback
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 reviewers found that the paper is interesting and easy to follow. As pointed out by a reviewer, the authors put effort into communicating the intuition. The authors performed experiments on six datasets, covering MRI, ultrasound, and fundus images, and performed different evaluations to show the advantages of the proposed method. On the other hand, the reviewers also raised concerns regarding the experimental details and the presentation of the results.
Reviewer #4 provided some useful comments. However, the AC disagrees with some of the weaknesses raised by the reviewer. For example, the reviewer considered proposing the Jacobian-based metric to be incremental. In contrast, the AC believes that establishing a simple yet functional metric for pre-training assessment, if successful, is an important contribution to the community.
Reviewer #4 also commented on the comparison models and the datasets. The AC considers some of these suggestions are infeasible for an 8-page MICCAI paper. As noted by other reviewers, although the paper is not flawless, it provides sufficient evidence that the proposed method is promising.
Overall, the paper meets the MICCAI standard in its current form.
