Despite the rapid advancement of 3D medical image segmentation foundation models trained on large-scale datasets, generalization to unseen data remains challenging. We observe that performance degradation is closely associated with localized voxel-level confidence distribution shifts between training and testing data. Existing Unsupervised Domain Adaptation (UDA) methods attempt to address sample-level domain discrepancies; however, they typically rely on computationally expensive gradient-based updates, which may lead to catastrophic forgetting and degradation of pretrained representations. To address these limitations, we propose a simple yet efficient post-hoc method termed Adaptive Cliff Alignment (ACA) to reduce voxel-level confidence distribution gaps in logit space. Our approach analyzes the shape of the predicted logit distributions and performs isotonic-regression-based quantile matching to align training and testing distributions. Importantly, the proposed alignment is confined to the model-uncertain region of the logit space rather than applied globally, enabling targeted correction while preserving confident predictions. Extensive experiments across multiple 3D medical image segmentation benchmarks demonstrate that ACA effectively mitigates train-test prediction distribution shifts and consistently improves segmentation performance in a train-free manner. Code is available at https://github.com/MelindaDong/aca-segmentation.
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4515_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
https://github.com/MelindaDong/aca-segmentation
N/A
@InProceedings{DonNan_ACA_MICCAI2026, author = { Dong, Nanyu AND Chen, Qi AND Verjans, Johan AND Liao, Zhibin}, title = { { ACA: Post-hoc Adaptive Logit Alignment for 3D Medical Image Segmentation } }, booktitle = {Medical Image Computing and Computer Assisted Intervention - MICCAI 2026}, year = {2026}, publisher = {Springer Nature Switzerland}, volume = {LNCS 16884}, month = {September}, page = {pending} }Review #1
- Please describe the contribution of the paper
The authors propose ACA, a method for post-hoc logit alignment for 3D segmentation methods, which help when there is a domain shift after adpatation.
- 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.
Here is a list of the major strengths of the paper.
- The motivation behind why current methods don’t work seems clear.
- While some important elements are missing (see below), the overall flow of the paper is easy to follow.
- The method, while simple, leads to an overall increase in performance. The analysis of the results also gives some interesting insights.
- 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.
Here are the major weaknesses, organised by category. 1.Hyperparameter choice. 1.1.The choice for using the range [-5, 10] for filtering is explained, but is not unique (e. g. , why 0.9999 and not 0.95). It would be nice to understand how these hyper parameters affect the final results? 1.2.Why do you choose 15% of the maximum gradient as the threshold (page 5)? Is this an important hyperparameter? How is the method affected by this?
2.Results 2.1.In Fig. 1 it seems that the distributions are well aligned (especially on the left). Is this before or after alignement? Is the shift in the test distribution somehow quantified? 2.2.How is the calibration of ACA compared to other methods? I’m concerned that the increased performance comes at the cost of a decrease in calibration.
3.Methods 3.1.You mention SAT is text-driven. How is it prompted exactly? 3.2.Since ACA adapts the logits class wise, is the same performed for other methods? 3.3.Part of the method is not entirely clear. How l and u and computed is clear. However, how these levels are used for the isotonic regression is not entirely clear. 3.4.Another related question is what is the point of thresholding at [-5, 10] if the method is only applied at [l, u]?
4.Others 4.1.I’m not sure I agree with your statement that “In segmentation tasks, overfitting often manifests as reduced predictive confidence near uncertain object boundaries in test samples. “. Modern models tend to be overconfident not underconfident ([1]). What are your arguments/sources here?
[1] On Calibration of Modern Neural Networks. Guo et al. 2017
- 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
Here is a list of some minor weaknesses: points that I think would help the overall quality of the paper if they were addressed.
- Some slight inconsistencies in the notations: x should be in bold for F(x) in the second paragraph on 3.1, z_train is used before it is introduced (and should be in bold)
- 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 the method is interesting and leads to good results. However, there are some important points which would need to be addressed, including questions regarding the hyperparameters, and some justifications for unclear points in the method.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
My initial rating was weak accept. Given that the authors have addressed most of my concerns in their rebuttal, I recommend an accept.
Review #2
- Please describe the contribution of the paper
This paper shows that 3D medical segmentation models often fail on new data because their voxel-level confidence distributions shift between training and testing. The authors introduce a simple post-hoc method called Adaptive Cliff Alignment (ACA). It aligns the uncertain logit region (in the paper described by the authors as “the cliff”) between train and test data using isotonic regression and quantile mapping, and there is noneed for retraining, labels, or gradients. The paper focuses only on the uncertain logits instead of the whole distribution, avoiding distortion of confident predictions. Experiments on multiple MSD datasets show consistent improvements over existing post-hoc and train-free approaches.
- 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 related work / state-of-the-art is addressed in an adequate way. Excellent and very helpful Fig. 2 illustrating the approach. A generally complex approach is explained in a consistent and precise way. Very impressive overall results on the MedDecathlon Dataset as shown in Table 1.
- 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.
Writing style is good, but there is no need to put relevant content into (). Either it is irrelevant - then delete - or it is relevant, then formulate proper sentences. Cf. “(to accommodate the fact that there is usually no available validation dataset)” Fig.1: blank missing before “Left”, cf. “MSD dataset [1].Left: two” Before Equation 1 reconsider the “,”, cf. “its probability density, i.e., FORMULA, ” First block of page 5 - many empiric values introduced, such as Gaussian sigma=8 and 15% threshold. Are these values generalizable for other domains? At least the Gaussian sigma is most probably related to the mentioned 300 bins only. Isotonic Quantile Mapping is an adequate strategy to align the logits. Abbrev. “i.i.d.” ok and familiar/common for statistics / ML specific papers - for MICCAI to please all reviewers this could be introduced the very first time. Discussion/Conclusion on page 8 highly relevant but quite excessively detailed description. Space still allows to do this but readability would be increased, if these highly important thoughts would be a bit compressed to highlight.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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?
Excellent writing style and highly impressive results for both, DSC and NSD. Proposed approach is very sound.
- 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 mainly format issues and writing style issues of R3 (me) were just ignored in the rebuttal - so unclear if these adaptions will make it to the final version. At least concerns regarding empiric Gauss sigma and so on were addressed at the begin.
Paper content is good - the writing style / layout need some improvement but the authors might be a bit unruly when it comes to suggestions.
Review #3
- Please describe the contribution of the paper
This paper proposes Adaptive Cliff Alignment, a post-hoc output logit calibration method for 3D medical image segmentation. The method identifies an “uncertainty region” in logit space and applies isotonic quantile mapping to align test logits to training logits, claiming improved generalization under distribution shift. Experiments on the Medical Segmentation Decathlon (MSD) show consistent but modest gains over several post-hoc 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.
It performs post-hoc logit alignment on the model output specifically in voxels with uncertainty. This makes the method practically useful in the medical imaging where fine-tuning is costly and can alter significantly the pretrained parameters and image annotation can be time-consuming and tedious.
- 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 independent binary formulation may lead to conflicting high-confidence predictions across multiple classes, and the proposed range filtering does not address such inconsistencies. 2) The paper positions the problem as unsupervised domain adaptation under distribution shift, however the experimental setup does not correspond to a standard DA scenario. The model (SAT) is pretrained on the Medical Segmentation Decathlon (MSD), and evaluation is also conducted within MSD across heterogeneous organ segmentation tasks. 3) While the method is computationally efficient at inference time, the reported improvements over strong baselines are relatively modest. In several datasets, improvements are marginal or comparable to simpler methods such as moment matching or optimal thresholding. All experiments are conducted on the MSD benchmark, which does not explicitly represent real-world clinical domain shifts such as, cross-hospital imaging variability or scanner-induced distribution shifts.
- 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
Several key components are introduced without sufficient theoretical or empirical justification, see fixed logit range filtering, Gaussian smoothing parameter, slope threshold for cliff detection, or N=1000 quantiles. The sensitivity of performance to these hyperparameters is not analyzed, raising concerns about robustness and reproducibility across datasets and model backbones.
- 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?
I recommend a weak reject due to concerns regarding novelty, experimental setting, and strength of claims. A major concern is the framing of the problem as domain adaptation, while the evaluation is conducted exclusively on the Medical Segmentation Decathlon (MSD). This benchmark primarily reflects cross-organ and cross-task variation within a single curated dataset family, rather than a true domain adaptation setting involving distinct imaging domains (e.g., cross-site, cross-scanner, or cross-institution shifts). As a result, the experimental setup does not convincingly support the claims of domain adaptation under realistic clinical distribution shifts. Furthermore, several key design choices are heuristic and lack sufficient ablation or theoretical justification.
- 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.
All 3 comments were clarified sufficiently. I would appreciate the written clarification in the manuscript that this methodology is for test-time logit alignment within the same dataset, and has not been applied for domain adaptation or different test-time datasets.
We thank all reviewers for their efforts and suggestions and will revise the paper accordingly. Code will be released on GitHub. We first address the common concerns:
Hyperparameter Sensitivity
Hyperparameters (#quantiles N=1000, #bins=300, Gaussian σ=8, gradient threshold=15%) were empirically selected through grid search. Performance remained stable under ~50% variations: #bins ∈ [200,400] and σ ∈ [6,10] changed DSC/NSD by <±0.3%, while N ∈ [500,2000] and threshold ∈ [0.1,0.3] changed DSC/NSD by <±0.1%.
Filtering Range
We select [-5,10] from l ∈ {-10, -5} and u ∈ {5,10}. Performance varies marginally (DSC/NSD <±0.35%), except for l=-10, u=10, which degrades performance. Quantile mapping requires collecting and sorting voxel-level logits from both train and test sets; Filtering reduces the number of collected voxels and hence lowers memory and sorting costs. The selected range excludes ~80% of voxels and achieves a balance between efficiency and performance.
Underconfidence near the boundary
The claim in [R2-Ref1] that “NN models tend to be overconfident” was established through comparisons between calibrated and uncalibrated predictions in classification settings, which we believe does not directly contradict our claim. We observe that test samples often exhibit lower prediction confidence near object boundaries, leading to under-segmented predictions despite correct localization. This behavior is illustrated in Fig. 2 (1st&3rd rows) and is consistent with prior work such as [arxiv.org/abs/2102.10365].
Domain shift in a clinical sense
We do not frame the problem as a clinical domain adaptation task. The terms “domain alignment” and “domain shift” were used to describe differences between the train and test logit distributions. Even under random train-test splits, deep learning models can overfit the train data, leading to performance degradation that results from train/test data differences. To avoid confusion, we will revise these terms with the more precise term “logit distribution alignment”.
Calibration
ACA results in <0.001% change in Brier score and ECE [R2-Ref1] compared to Moment Matching (MM) and SAT baseline. While we used the term “calibration”, our objective is to improve generalization by mitigating differences between train and test logit distributions, rather than calibrating prediction confidence to its true probabilistic interpretation. To avoid confusion, we will remove the term “calibration” from the paper.
R2
2.1.Fig.1 shows distributions before alignment. Distribution shift can be quantified by the overlapping coefficient (OVL). The left panel (liver) shows high overlap (OVL=0.96), likely because the liver is a large, well-defined structure in CT.
3.1-3.SAT is prompted with just the class names, e.g., ’anterior hippocampus’; All methods are applied on logits class-wise for fair comparison; We subdivide 1000 quantiles from l to u for isotonic quantile mapping (see the 2nd paragraph under Isotonic Quantile Mapping).
R4
1.SAT’s text prompt-based segmentation adopts a binary formulation, where each prompt targets a single class. Although multiple classes may yield high-confidence predictions simultaneously, this does not affect the proposed range filtering, which is applied independently to class-wise logit distributions. Post-hoc softmax normalization could be used for final class selection but would require additional architectural changes and fine-tuning of SAT, which is beyond the scope of this work.
3.The average improvement of ACA over the 2nd-best method, MM, is +0.96% DSC and +0.82% NSD. Notably, MM itself only improves over the baseline SAT model by +0.25% DSC and +0.18% NSD. Since all compared methods are training-free, ACA provides a meaningful, large gain over the strong SAT foundation model baseline. Moreover, the improvements are more pronounced in lesion-related tasks (on average +2.08% DSC and +1.85% NSD over the baseline).
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.
Reviewers are split: one recommends strong accept, one recommends weak accept, and one recommends weak reject. After careful consideration, I recommend sending this paper for rebuttal. The paper addresses a practically important problem — post-hoc logit calibration for 3D medical image segmentation under distribution shift — and the core idea of focusing alignment on the uncertain “cliff” region is well-motivated and intuitive. The method requires no retraining, labels, or gradients, which is a significant practical advantage. Results on MSD are generally strong and the paper is clearly written. However, several concerns need to be addressed. Multiple key hyperparameters — the logit range [-5, 10], the 15% gradient threshold, the Gaussian smoothing parameter, and N=1000 quantiles — are introduced without sufficient justification or sensitivity analysis, raising reproducibility concerns. The experimental setup does not fully correspond to a standard domain adaptation scenario, as evaluation is conducted within MSD across tasks rather than across genuinely distinct clinical domains. The claim that segmentation models are underconfident near boundaries contradicts established findings in calibration literature. The relationship between calibration and the reported performance gains is not analyzed.
The authors are asked to clarify the following points in their rebuttal:
- The theoretical or empirical basis for the hyperparameter choices, based on existing results or analysis already in the paper.
- The justification for the underconfidence claim near boundaries, with reference to existing literature or experiments.
- Whether the experimental setup genuinely represents distribution shift in the clinical sense, and how the authors would characterize the nature of the shift being addressed.
- Whether existing results in the paper support claims about calibration behavior.
- 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.
Following the rebuttal, all reviewers reached a consensus to accept the paper. The authors are encouraged to refine their claim regarding domain alignment and address the writing style concerns raised by Reviewer 3 in the final version to prevent potential misunderstandings.
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.
All reviewers confirmed that most of their concerns are addressed. However, minor issues remain. Please try to have that addressed in the camera-ready version.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
I recommend accept: post-rebuttal unanimous Accept.
back to top
