ACA: Post-hoc Adaptive Logit Alignment for 3D Medical Image Segmentation

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

Review #2

Review #3

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

Meta-review #2

Meta-review #3

back to top

Link nội dung: https://chodichvu.vn/hoc-3d-a91041.html