acceptodds
Under review as a conference paper at ICLR 2027

Source-Free Test-Time Adaptation for Brain Tumor Segmentation by Jointly Handling Covariate and Label Shifts

Abstract

In real deployment scenarios, covariate and label shifts typically co-occur, yet most Test-Time Adaptation (TTA) methods account only for the covariate shift. This limitation arises primarily because covariate shift adaptation assumes a stationary label distribution, whereas label shift correction presumes a fixed class-conditional input distribution. Consequently, neither approach is directly applicable when both shifts are simultaneously present. Moreover, the label shift is inferred from model predictions, while the covariate shift distorts these predictions, making accurate label shift correction feasible only after the covariate shift has been effectively mitigated. We further observe that the label shift is particularly detrimental in brain tumor segmentation, where the evaluation regions are hierarchically nested and any deviation in a class propagates to every region containing that class, thereby amplifying the shift in the predictions toward the source-domain label distribution. We term this effect *hierarchical label shift*. Based on the above analysis, we propose **H**ierarchical-**R**estoring **A**daptation and **C**orrection (HRAC), a source-free TTA framework that jointly handles the covariate shift and the hierarchical label shift in an *adapt-then-correct* manner by explicitly structuring adaptation and correction around the nested evaluation regions. Our HRAC comprises two synergistic components: (i) **S**tructure-**C**onsistency **A**daptation (SCA), which mitigates the covariate shift by restoring the spatial structure it corrupts, with consistency losses imposed across the nested regions; and (ii) **H**ierarchical **L**abel **S**hift **C**orrection (HLSC), which corrects the hierarchical label shift by applying a bounded logit bias to the violated region ratios. To the best of our knowledge, HRAC is the first source-free method to jointly mitigate the covariate shift and correct the hierarchical label shift without estimating a target prior. Extensive experiments on three BraTS benchmarks show that HRAC outperforms seven state-of-the-art adaptation baselines in mean Dice, mean HD95, and mean IoU, demonstrating its effectiveness and strong generalization in brain tumor segmentation.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Test-Time Adaptation, Brain Tumor Segmentation, Label Shift, Covariate Shift

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_YniyOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Hierarchical-Restoring Adaptation and Correction (HRAC) for source-free test-time adaptation in brain tumor segmentation under changes in image appearance and tumor composition. HRAC comprises two stages. Structure-Consistency Adaptation (SCA) updates BatchNorm affine parameters using flip consistency and a partial-order consistency loss defined over the nested enhancing tumor (ET), tumor core (TC), and whole tumor (WT) regions. Hierarchical Label Shift Correction (HLSC) subsequently applies per-volume additive logit biases based on violations of fixed bounds on intra-tumor composition ratios, without explicitly estimating the target class prior. The main evaluation transfers a model trained on BraTS 2023 adult glioma to three pediatric or Sub-Saharan African cohorts, comparing against seven adaptation/correction methods. HRAC achieves the best reported mean Dice, HD95, and IoU across these targets.

Critical strengths
  1. The paper attempts to address a meaningful limitation of deploying adult-trained segmentation models on populations with different tumor compositions.

  2. HRAC improves computational efficiency by restricting optimization to BatchNorm affine parameters, requiring only two adaptation steps per volume, and applying the final correction without gradient-based optimization.

  3. The SCA component achieves a large performance gains across all three segmentation targets.

Critical weaknesses
  1. Insufficient justification of the thresholds. The choice of the structural composition prior (0.4, 0.6, 0.6) lacks a clear derivation or supporting references.

  2. The SCA component contributes the largest performance gain. However, enforcing consistency between original and augmented views is a common strategy in TTA methods, and the manuscript does not sufficiently clarify the novelty of SCA relative to existing approaches.

  3. Limited and uneven benefits of HLSC. Despite being a central contribution, HLSC consistently reduces WT and ET Dice in Tables 2 and 3, while yielding only marginal gains in mean Dice.

  4. Limited comparisons with recent methods. The evaluation does not include any methods proposed in 2026, making it difficult to assess HRAC’s competitiveness against recent advances in TTA.

  5. Please include comparisons with additional medical TTA methods, particularly GraTa [1], which is also designed for medical image segmentation and employs a consistency constraint between original and augmented views. [1] Chen Z, Ye Y, Pan Y, et al. Gradient alignment improves test-time adaptation for medical image segmentation[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2025, 39(3): 2429-2437.

Questions for authors
  1. How were the thresholds (0.4, 0.6, and 0.6) selected, and what empirical evidence supports their validity?

  2. How does SCA differ from existing TTA methods that enforce consistency between original and augmented views?

  3. Why does HLSC consistently lead to lower performance for WT and ET?

  4. Could the authors include comparisons with more recent medical TTA methods published in 2026, or clarify why these methods were omitted?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Yniy (1/1)

We greatly appreciate your constructive and insightful comments. We address all weaknesses below and will incorporate the corresponding changes into the revised manuscript.


W1&Q1: Insufficient justification of the thresholds.

The three values have a clear empirical basis, but they are used as one-sided structural bounds rather than as estimates of the target composition, so no derivation to an exact value is required.

(1) What the three values are. ρ_TC = 0.4 is a floor on the core share of the whole tumor, ρ_ED = 0.6 is a ceiling on the edema share, and ρ_ET = 0.6 is a floor on the ET share of the tumor core. The measured ratios are rTC=PTC/PWTr_{TC} = P_{TC} / P_{WT}, rED=PED/PWTr_{ED} = P_{ED} / P_{WT}, and rET=PET/PTCr_{ET} = P_{ET} / P_{TC}. Fig. 1(a) shows that in adult glioma the core makes up about four tenths of the whole tumor, edema about six tenths, and ET about six tenths of the core. Taking these proportions to one decimal place gives (0.4, 0.6, 0.6). They are coarse prior knowledge about typical composition rather than statistics estimated from source training labels. They are fixed before deployment, and adaptation accesses neither source data nor target labels, so the source-free setting is preserved.

(2) Why exact values are unnecessary. The constraints are one-sided. The bound decides whether a direction needs correction and the measured deviation decides by how much. The correction is zero when no bound is crossed, so a bound only needs the right scale and direction rather than an exact match to the typical composition of any cohort. We treat 0.4 as a loose floor that fires only when the core is clearly under-predicted.

(3) Experimental verification. A 7 × 7 × 7 sweep of ρ_s across the three dimensions (343 configurations, Appendix C.2 and Figure 4) changes Mean Dice only between 60.19 and 61.93, with every configuration within 0.90 points of the deployed one. The same bounds are shared across BraTS 2023 PED, BraTS 2025 PED and SSA without per-population adjustment, and HRAC achieves the best mean Dice and the best mean HD95 on all three (Table 1). What the method relies on is the directional constraint of the bounds, not any precise threshold.


W2&Q2: The novelty of SCA relative to view-consistency methods.

The difference between SCA and existing consistency methods lies in what the consistency is used for. In existing methods it serves as a smoothing regularizer or, as in GraTa, as the pseudo gradient, with the entropy loss supplying the auxiliary one for alignment. In SCA the flip runs first, making the prediction more confident and structurally self-consistent, so that a disagreement with its mirrored copy reflects a misplaced structure rather than noise. On top of that confident prediction the poset term takes over and suppresses the responses still falling outside the tumor body. What SCA recovers is where the structure sits and how far it extends, not a smoother probability map.


W3&Q3: Why do WT and ET decrease?

The gain of HLSC concentrates on TC, and it is large. Adding HLSC on top of SCA gains 7.78 in TC Dice, together with 1.05 in mean Dice and 15.62 mm in HD95. TC is exactly where the population shift acts and the weakest region for the source model (32.88 Dice).

The WT and ET decreases are the price of this concentration, and neither indicates a failure of the correction. The WT decrease is slight. WT always equals TC plus ED, and rTC+rED≡1r_{TC} + r_{ED} ≡ 1 holds in all cases. Reallocation between core and edema does not change total foreground volume, so the correction leaves WT nearly in place, and with the source WT already close to GT, there is little room to gain. The ET decrease is a side effect of the TC correction, not the objective of the ET term. The correction uses rET=PET/PTCr_{ET} = P_{ET} / P_{TC}, whose denominator is the inflated TC, and raising NCR compresses rETr_{ET}, which triggers the ET term and raises ET predictions. Experimentally, the resulting false positives arise almost exclusively in cases that should contain no ET.


W4-5&Q4:Comparison with recent methods.

GraTa[1] and LATTA[2] have been added, reproduced under our protocol, and compared on BraTS 2023 PED. With the most recent 2025 and 2026 methods included, HRAC still leads in mean Dice and HD95.

MethodWTTCETMean DiceMean HD95
GraTa (AAAI 2025)82.5739.1455.0158.9051.33
LATTA (WACV 2026)81.8736.7354.4157.6758.78
HRAC (Ours)82.1950.4051.8761.4929.61

[1] Chen Z, Ye Y, Pan Y, et al. Gradient alignment improves test-time adaptation for medical image segmentation. AAAI 2025.

[2] Weerasinghe, Thilina, et al. Logit-Adjusted Test-Time Adaptation under Partial Class Imbalance. WACV 2026.

Thank you for the detailed response and additional experiments. My main concerns remain. For example, the explanation of SCA clarifies the intended role of flip consistency and the partial-order term, but it does not yet establish which part of the gain comes from their proposed interaction, as opposed to established consistency-based adaptation. A controlled comparison isolating that interaction would make the novelty claim more convincing. Overall, the rebuttal addresses some of my questions, but the remaining concerns do not change my assessment. I will maintain my original rating.

AuthorsOpenReview ↗

Response to Reviewer Yniy

We thank you for the continued engagement. To clarify the controlled comparison you requested, we have assembled a decomposition from the ablation tables, paired statistics, and a new controlled run.

As summarized below, the first table shows the cumulative gain of adding one component at a time, while the second table presents the ablation results where the partial-order term is replaced by the established entropy minimization

Table 1. Cumulative module addition

Configuration (Base: Source model)Mean DiceΔ
Source model50.81—
+ Flip consistency59.66+8.85 (vs. Source)
+ Poset (SCA)60.43+0.77 (vs. +Flip)
+ HLSC (HRAC) [Proposed]61.49+1.05 (vs. SCA)

Table 2. Controlled comparison (Entropy replacement)

Configuration (Base: Source model)Mean DiceΔ
Source model50.81—
+ Flip + Entropy (no HLSC)59.60−0.83 (vs. SCA)
+ Flip + Entropy + HLSC60.78−0.71 (vs. HRAC)

We state plainly that flip consistency alone (the established component) accounts for 8.85 of the 9.62-point improvement over the unadapted baseline. However, neither the partial-order term alone (59.37) nor HLSC alone (51.87) recovers the full result.

Entropy minimization does not extend the established component (59.60 vs. 59.66). The partial-order term outperforms it by 1.36 in TC Dice , with identical WT. The critical difference lies in where the penalty is applied. The partial-order term penalizes only responses outside the tumor body, whereas entropy minimization sharpens every uncertain voxel, including correct ones inside the core. These outside responses are exactly the failure mode that consistency regularization fails to address. The same replacement inside the full pipeline yields the same conclusion (60.78 vs. 61.49).

The specific contribution of HLSC is evident in the TC region, which is exactly where the population shift distorts class composition. Adding HLSC improves the TC Dice by 7.78 and the mean HD95 by 15.62 mm.

Therefore, our novelty claim rests not on the consistency component itself, but on the structural constraints (SCA and HLSC), which repair the specific failure mode that established consistency-based adaptation cannot. We will incorporate this controlled decomposition explicitly into the revised manuscript so the source of the gain is verifiable at a glance.

We respectfully ask you to reconsider the novelty assessment in light of this evidence, and we hope this clarification is useful to the AC as well.

Reviewer_B2nuOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HRAC, a source-free test-time adaptation framework for brain tumor segmentation that aims to address both covariate shifts and label shifts between source and target domains. The method consists of two main components: Structure Consistency Adaptation (SCA) and Hierarchical Label Shift Correction (HLSC).

SCA adapts the segmentation model by combining a partial-order consistency (Poset) loss, designed to encourage anatomical relationships between nested tumor regions (ET⊆TC⊆WTET \subseteq TC \subseteq WT), with a consistency loss between predictions from original and horizontally flipped MRI volumes. During adaptation, only the BatchNorm affine parameters are updated. HLSC subsequently adjusts the segmentation logits using volume-level estimates of tumor-region proportions and predefined structural bounds, aiming to correct potential label shifts without access to source data.

The paper evaluates HRAC using a 3D U-Net trained on BraTS 2023 GLI and tested on pediatric and sub-Saharan African brain tumor datasets. The experimental results show that HRAC generally improves segmentation performance over the source-only model and several existing test-time adaptation methods. Ablation studies investigate the contributions of SCA and HLSC, with SCA providing most of the improvement in mean Dice and HLSC offering additional benefits, particularly in tumor-core segmentation and boundary accuracy.

Critical strengths
  • The paper addresses the simultaneous presence of covariate and label shifts in brain tumor segmentation, which is a relevant challenge for deploying medical image segmentation models across different patient populations and acquisition settings.

  • The paper highlights how changes in the proportions of individual tumor classes affect the nested evaluation regions (ET, TC, and WT). This provides an interesting perspective on label shift in brain tumor segmentation.

  • Evaluation across multiple target domains: The authors evaluate HRAC on three BraTS datasets representing different patient populations, including pediatric and sub-Saharan African cohorts, and compare it against seven existing adaptation methods.

  • The proposed method achieves improvements over the source-only model and competitive TTA baselines.
Critical weaknesses

1- Unclear motivation and interpretation of the partial-order consistency loss

  • The motivation for the proposed partial-order consistency (Poset) loss is not fully clear to me. The authors argue that the loss enforces the anatomical hierarchy ET⊆TC⊆WTET \subseteq TC \subseteq WT. However, this hierarchy is already guaranteed by the model’s output representation. Specifically, the model predicts four mutually exclusive classes (BG,ED,NCR,ETBG, ED, NCR, ET), and the tumor regions are constructed as:

pTC=pNCR+pET,pWT=pED+pNCR+pETp_{TC}=p_{NCR}+p_{ET}, \qquad p_{WT}=p_{ED}+p_{NCR}+p_{ET}.

Consequently, pET≤pTC≤pWTp_{ET}\leq p_{TC}\leq p_{WT} always holds. For example, if the model assigns a probability of 0.8 to ETET and 0.1 to NCRNCR at a voxel, the corresponding TCTC probability is automatically 0.9. Therefore, the model cannot predict high ETET confidence while simultaneously predicting low TCTC confidence, which appears to be the type of hierarchical inconsistency that the proposed loss aims to prevent.

It is unclear to me what anatomical inconsistency is the Poset loss actually correcting if the network architecture already guarantees the proposed hierarchy?

  • Given that the hierarchical relationships are already guaranteed by construction, it is unclear what the proposed Poset loss actually encourages during optimization. In particular, the loss can produce nonzero penalties even when the hierarchy is perfectly satisfied. Consider the first term in Eq. 5:

MET(1−M~TC)M_{ET}(1-\widetilde M_{TC}).

Suppose that MET=MTC=0.5M_{ET}=M_{TC}=0.5, and that dilation does not change the TCTC mask. The contribution of this voxel to the numerator of the first term is 0.5(1−0.5)=0.250.5(1-0.5)=0.25. In contrast, if both values are 0.90.9, this contribution decreases to 0.9(1−0.9)=0.090.9(1-0.9)=0.09. Both examples satisfy the required hierarchical relationship, yet the more confident configuration produces a smaller penalty contribution.

This distinction is important because the difference in penalties does not reflect a difference in anatomical consistency: neither prediction violates the hierarchy. This suggests that the proposed loss may function as a form of confidence-based and spatial regularization rather than explicitly correcting hierarchical inconsistencies. In particular, it appears to encourage certain confidence configurations of the nested regions.

I therefore encourage the authors to clarify the actual mechanism of the Poset loss and compare it against simpler confidence-based regularization objectives, particularly entropy minimization. A comparison between Flip + Poset and Flip + Entropy Minimization, under identical adaptation settings, would help determine whether the proposed formulation provides benefits beyond conventional confidence regularization. An additional ablation without morphological dilation would help isolate the contribution of spatial regularization.

2-Limited and inconsistent contribution of HLSC

The contribution of the proposed HLSC component is relatively limited according to the results in Table 2. While SCA improves mean Dice substantially, adding HLSC provide relatively small additional improvement. Given the relatively small improvement in mean Dice, it is unclear whether the overall improvement is statistically significant or consistently observed across patients.

Additionally, the effect of HLSC varies across tumor regions. It increases the Dice score in some regions and decreases in some other. I would also be interested in understanding what leads to this inconsistent behaviour.

3-Missing information about source model data augmentation

The paper trains a 3D U-Net on BraTS 2023 GLI, but the source training description does not clearly specify whether data augmentation is used. This is particularly important because the paper addresses domain shifts. Strong intensity and spatial augmentations during source training have been shown to substantially improve robustness to acquisition and population differences in both earlier [1] and more recent [2] works.

Furthermore, nnU-Net is one of the most widely adopted frameworks for medical image segmentation, making it an important baseline for evaluating the practical effectiveness of the proposed method. I would therefore recommend training an nnU-Net model on the source domain using the augmentation framework introduced in [2] and evaluating whether HRAC still provides substantial improvements under the same domain shifts.

[1] Zhang et al., Generalizing Deep Learning for Medical Image Segmentation to Unseen Domains via Deep Stacked Transformation. TMI 2021

[2] Molinier and Moller et al., One Sequence to Segment Them All: Efficient Data Augmentation for CT and MRI Cross-Domain 3D Spine Segmentation, MICCAI 2026

4-Gradient update steps for test-time adaptation

The paper explicitly states that HRAC performs two gradient updates per target volume. However, it is unclear whether the competing TTA methods were evaluated using the same number of gradient updates. Since some TTA methods are sensitive to the number of adaptation steps, I wonder whether differences in adaptation budgets could influence the reported comparisons.

Additionally, I would be interested in understanding how HRAC performs with more adaptation steps. Some TTA methods initially improve with additional updates but eventually experience performance degradation due to model drift. Does HRAC exhibit similar behavior, or does the proposed approach provide more stable adaptation? An ablation evaluating different numbers of gradient updates per target volume would help clarify this.

5-Limited investigation of alternative consistency transformation

SCA uses horizontal flipping as transformation for the consistency loss. However, I am not sure why this was chosen as the transformation instead of other transformations such as intensity and contrast which are particularly relevant to MRI domain shifts. The authors could compare flip consistency with other transformations (or perhaps their combinations) to justify and show the effectiveness of this choice.

6-Comparison with Jang & Chung (2025) and Park et al. (2023) The paper discusses that Jang & Chung (2025) and Park et al. (2023) as existing approaches that address covariate and label shifts simultaneously. However, neither method is included in the experimental comparisons. Given that jointly handling these two shifts is a central motivation of the paper, I would be interested in understanding why these methods were excluded.

Minor comment

  • Figure 2(a) appears inconsistent with the methodology. The figure seems to illustrate T1 and T2 modalities as separate branches, whereas the text describes consistency between the original MRI volume and its horizontally flipped version.
Questions for authors

1.1-Given that the model’s output representation already guarantees pET≤pTC≤pWTp_{ET} \leq p_{TC} \leq p_{WT}, what specific anatomical inconsistencies is the proposed Poset loss intended to correct?

1.2-Could the authors further explain its optimization behavior and clarify how it differs from conventional confidence regularization? In particular, could they compare Flip + Poset against Flip + Entropy Minimization and investigate the contribution of morphological dilation?

2.1-Table 2 shows that adding HLSC to SCA improves mean Dice only marginally (60.43% to 61.49%). Could the authors paired statistical significance tests to establish the reliability of these improvements?

2.2-Why does HLSC improve TC Dice while decreasing WT and ET Dice?

3.1-Did the authors use data augmentation when training the source model? If so, could they provide details about the augmentation strategy?

3.2-Could the authors evaluate HRAC using a stronger source model trained with nnU-Net, ideally incorporating the augmentation framework proposed in [2], to demonstrate its effectiveness with a widely adopted medical image segmentation pipeline?

4.1-Were the competing TTA methods evaluated using the same number of gradient updates per target volume as HRAC?

4.2-Could the authors provide an ablation investigating how HRAC’s performance changes with increasing numbers of adaptation steps? In particular, does performance remain stable, or does it eventually degrade due to model drift?

5-What motivated the choice of horizontal flipping for the consistency loss? Have the authors considered intensity- or contrast-based transformations, which may be more directly relevant to MRI domain shifts? Could they compare these alternatives to justify the current design choice?

6-Why Jang & Chung (2025) and Park et al. (2023) were excluded from the comparisons?

3 comments
AuthorsOpenReview ↗

Response to Reviewer B2nu (1/3)

We greatly appreciate your constructive and insightful comments. We address all weaknesses below and will incorporate the corresponding changes into the revised manuscript.


W1&Q1.1-1.2. What the Poset loss corrects, and how it differs from confidence regularization.

We thank the reviewer for this derivation. The hierarchy argument holds unconditionally, since four mutually exclusive output classes make pET≤pTC≤pWTp_{ET}≤p_{TC}≤p_{WT} always hold. We no longer claim that this loss enforces the hierarchy, and we will revise the corresponding description in the revision.

(1) Design motivation for Poset? Its intent is to suppress predictions outside the tumor core. Enhancing tumor lies within the core. High-confidence ET responses outside the core are mostly false positives. The region nesting does not constrain where a response occurs, so such responses need a dedicated penalty.

(2) What does it actually correct? It corrects responses outside the tumor core, not the hierarchy. The hierarchy is guaranteed by construction. Four mutually exclusive output classes make pET≤pTC≤pWTp_{ET}≤p_{TC}≤p_{WT} always hold, and among the 135,508 predicted ET voxels, zero violate ET⊆TC⊆WTET⊆TC⊆WT. No inconsistency in the hierarchy therefore exists for this term to correct. Without dilation the term responds only to confidence, not to location. With the dilation gate, the in-core contribution is zeroed, removing 89% of the penalty mass, and what remains depends only on whether a voxel lies outside the core. The mechanism is spatial under dilation and confidence-shaped only without it.

(3) How does it differ from confidence regularization? At equal strength the two differ only in where the penalty is applied. Against entropy minimization (λ=1) the Poset term gains 0.71 Dice (p = 1.3e-6, 72 of 99 cases improved) and reduces HD95 by 4.55 mm (pp = 0.022). At matched magnitude, entropy minimization (λ=100) sits 3.21 points below the no-term baseline (57.67 versus 60.88) while the Poset term stands 0.61 above it (61.49 versus 60.88). Entropy minimization sharpens every uncertain voxel, including correct ones inside the core, whereas the dilation gate confines the Poset penalty to outside-core responses. Applied only outside the core, the same weight helps. Applied everywhere, it hurts. Without the gate the term is a pure confidence penalty, and it still gains 0.45 Dice and 3.31 mm HD95, with the gain saturating for k≥5 as the neighborhood covers nearly all true ET responses.

ConfigurationMean DiceMean HD95Magnitude
No confidence term (λposet\lambda_{poset}) = 0)60.8834.14—
poset k=1 (reduces to the confidence penalty)61.3330.830.1142
poset k=3 (deployed)61.4929.610.01258
poset k=561.6329.57—
poset k=761.6329.56—
Entropy minimization λ\lambda=160.7834.16~1.5e-3
Entropy minimization λ\lambda=10057.6749.46comparable to poset

W2&Q2.1-2.2. Limited and inconsistent contribution of HLSC.

(1) Validity of HLSC results. Per-case paired tests of HRAC against SCA-only (n = 99) show that adding HLSC gains 1.05 in mean Dice (p = 4.3e-6, 73 of 99 cases improved), 7.78 in TC Dice (p = 8.0e-15, 83 of 99 cases improved), and reduces HD95 by 15.62 mm (p = 1.2e-6, 68 of 99 cases improved). The gain concentrates on the tumor core, the region whose composition the population source distorts, and it comes from a single gradient-free logit correction that adds no iteration cost.

Metric (HRAC vs SCA-only)Δ\Deltapimproved/worsened
Mean Dice+1.054.3e-673/26
Mean HD95-15.621.2e-668/31
TC+7.788.0e-1583/16
WT-0.850.03540/59
ET-3.760.01224/75

(2) Why do WT and ET decrease? Neither decrease indicates a failure of the correction. The WT decrease is slight. WT always equals TC plus ED, and rTC+rED≡1r_{TC} + r_{ED} ≡ 1 holds in all cases. Reallocation between core and edema does not change total foreground volume, so the correction leaves WT nearly in place, and with the source WT already close to GT (predicted-to-GT volume ratio of about 0.90), there is little room to gain. The ET decrease is a side effect of the TC correction, not the objective of the ET term. The correction uses rET=PET/PTCr_{ET} = P_{ET} / P_{TC}, whose denominator is the inflated TC, and raising NCR compresses rETr_{ET}, which triggers the ET term and raises ET predictions. Experimentally, the resulting false positives arise almost exclusively in cases that should contain no ET.

AuthorsOpenReview ↗

Response to Reviewer B2nu (2/3)

W3&Q3.1-3.2. Missing information about source model data augmentation.

The 3D U-Net source model was trained without any data augmentation, and the revision will state this explicitly. We retrained it with the nnU-Net v2 strong-augmentation pipeline and re-ran the full comparison under the same protocol. HRAC still achieves the best mean Dice and the best HD95.

(1) What does the re-evaluation show? We retrained the GLI source model with the full strong augmentation pipeline of nnU-Net v2 (Gaussian noise and blur, brightness 0.75 to 1.25, contrast, low-resolution simulation, gamma 0.7 to 1.5, mirroring with rotation, scaling, and elastic transforms), which spans both the intensity and spatial augmentation families, and re-ran the full table on BraTS 2023 PED (n = 99).

Method (PED2023, nnU-Net)WTTCETMean DiceMean HD95
w/o adaptation86.3047.5553.3762.4134.38
Tent86.6247.2853.5962.5033.84
EATA86.7446.5353.5562.2733.65
SAR86.7047.0753.5862.4533.63
TEGDA86.6646.9453.5962.3932.47
Undoing UDA86.4547.5753.4562.4934.06
BLDA28.0543.9841.1237.7267.73
HRAC (Ours)86.9550.5153.6263.6927.96

SmaRT is not listed because its implementation does not run on the nnU-Net pipeline.

Under this stronger source model, HRAC remains best. It gains 1.28 in mean Dice (p < 0.001, 73 of 99 cases improved) and 2.96 in TC Dice (p < 0.001, 81 of 99 cases improved), leads the strongest baseline Undoing UDA by 2.94 in mean Dice (p < 0.001), and reduces HD95 by 6.42 mm (p < 0.001, 71 of 99 cases improved), with the best HD95 in all three regions, the largest on ET (62.49 to 47.45 mm). WT (+0.65, p = 0.162) and ET (+0.25, p = 0.177) remain on par with the strong source model, and the share of cases with empty ET predictions drops from 13.1% to 9.1%.

(2) A note on this backbone. Entropy-based methods do not improve it (|Δ| ≤ 0.14), SCA-only (62.39) sits at the source level (62.41), and the gain of the full method comes almost entirely from the composition correction. InstanceNorm has no running statistics to recalibrate and the prediction entropy is only 5e-4 (6e-2 on the 3D U-Net), so consistency-based adaptation has little to act on, which points to the normalization type and the uncertainty scale rather than the architecture.


W4&Q4.1-4.2. Gradient update steps for test-time adaptation.

(1) Are the comparisons budget-matched? The step budgets, counted as backward passes per target volume, are Tent 2, EATA 1, SAR 1, TEGDA 1, SmaRT 1, and HRAC 2. Re-running the baselines aligned at 2 steps leaves the conclusion unchanged. HRAC (61.49) still leads TEGDA (58.87) by 2.62 points. Tent already runs at the same step count and learning rate as HRAC (2 steps), and its gain over the source model is +6.28 against +10.68 for HRAC. Aligning the budgets therefore does not change the comparison, and a budget-matched baseline does not reach the gain of HRAC.

Method1 step2 steps
EATA57.5457.77 (+0.23)
SAR57.3957.19 (-0.20)
TEGDA58.8958.87 (-0.02)
SmaRT55.8155.97 (+0.16)

(2) Does HRAC degrade with more steps? With all other settings fixed, Dice first rises and then crashes (table below). HRAC does not degrade immediately, it improves monotonically up to 8 steps. The 16-step collapse has a direct cause: accumulated cross-volume updates saturate the soft masks and thereby shut off the structural signal. The affine parameters persist across the stream, so 16 updates per volume accumulate roughly 8 times more updates than the deployed setting, and the persistent updates push the soft region masks toward saturation. The poset term is an M(1−M~)M(1 − \tilde{M})-type quantity, where M~\tilde{M} is the dilated container mask, and it vanishes when the masks saturate. Under the continuous setting without per-volume reset, LposetL_{poset} drops to about 1e-6 from step 4 onward, so the structural constraint is lost. From that point the updates are driven by the model's own predictions and errors accumulate along the stream, with the last 20 cases scoring 25.17 against 67.61 for the first 20. The control with a state reset at every volume confirms this attribution: with the reset, LposetL_{poset} stays at 0.15 to 0.21 and the cumulative degradation disappears, so the collapse comes from cross-volume accumulation rather than from the loss itself.

StepsMean DiceMean HD95
160.7232.08
261.4929.61
462.6526.85
863.8224.96
1653.7039.56
AuthorsOpenReview ↗

Response to Reviewer B2nu (3/3)

W5&Q5. Why flip, not appearance transforms.

SCA repairs a spatial failure through a consistency constraint between two views, and that constraint can only act on spatial structure if the two views place voxels at different locations. The flip forms a pointwise mirror correspondence between them, and consistency is required to hold pointwise between mirrored positions, which forces the model to place the structure back where it belongs. Appearance transforms change no voxel position, their consistency reflects intensity sensitivity only, and they cannot constrain this failure mode.

TransformMean DiceΔ\Delta vs flipHD95
flip (Ours)61.49—29.61
gamma61.05-0.4435.43
contrast60.90-0.5930.99
bias60.69-0.8035.48
flip+gamma61.72+0.2429.51

The comparison confirms this. Among the single transforms flip is the highest at 61.49, and the three appearance transforms are all lower (61.05, 60.90, and 60.69), because they do not constrain the spatial failure. Flip+gamma adds 0.24 on top (61.72), and the appearance transform contributes a constraint on robustness to intensity variation while leaving the spatial one untouched, so the two are complementary.


W6&Q6. Comparison with Jang & Chung (2025) and Park et al. (2023).

Neither Jang & Chung (2025) nor Park et al. (2023) has a public implementation, and neither is source-free: Park's adapters rely on source-domain information, and the prediction refinement module of Jang & Chung (2025) is trained with labeled source data. Running them under our protocol would require a reimplementation from the paper, and the fidelity of that reimplementation cannot be verified, which would make the comparison unreliable. They are therefore not included as comparison methods in our tables.


M1. Do the two branches use different modalities?

The two branches take the same four-channel volume. All four sequences form a single input, the upper branch processes it, and the lower branch processes its flipped copy, with shared weights. The arrows drawn from individual modality rows created the impression of separate modality branches, and we will revise the figure in the revision to remove this ambiguity.

Reviewer_4xBEOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

A source-free test-time adaptation method is developed for brain tumor segmentation. The authors consider the covariate shift and label shift simultaneously. Structure consistency adaptation (SCA) is used to deal with covariate shift , and hierarchical label shift correction (HLSC) is proposed for label shift. The brats datasets were used for experiments, and the authors considered multiple target domains for evaluation.

Critical strengths

1, Dealing with covariate shift and label shift is a practical setting, where the test domain images may have not only image appearance shift but also structural shift. Most existing methods focus on the image appearance shift while ignoring the structural shift (label distribution).

2, The proposed has a good motivation, where SCA is leveraged for dealing with the covariate shift, and HLSC is used for label distribution shift.

3, Multiple testing domains were used for experiments, and the authors' performance obtained the best results in most cases.

Critical weaknesses

1, The covariate shift and label distribution shift are not well explained in the introduction. The reasons for these two types of shift should be described more clearly.

2, The structure consistency adaptation is too simple and has a limited novelty. Encouraging prediction consistency under image transforms has been widely used as regularization in annotation-efficient learning methods. Though using such regularization for domain adaptation where the target label is unavailable is reasonable, such a method does not introduce new information.

3, Using the prior information of the ratio of different structures is not new as well. The authors have some over-claim about the novelty of this work. The label shift correction has been studied in previous works, the authors claim that "none of them handles the covariate shift and the label shift at once", but they just used two simple and existing techniques to deal with these problems respectively. Each of them are established methods, and the new knowledge provided by this work is marginal.

4, The results are not convincing enough. Table 1 shows that in many cases, the proposed method did not outperform the existing ones, especially on the ET and WT class. The authors also did not provide statistical test results to demonstrate if the improvement from the existing method is significant or not.

5, Important details about the experiments are not provided. The authors deal with test time adaptation, but did not mention how to deal with the 3D volumes for inference. The batch size, patch size and number of iterations were not provided for model update before inference for the final output on each testing case. Without these details, it is not clear if the comparison is fair.

Questions for authors

1, In the label shift setting, it is not possible to know the distribution of labels in the target domain when no ground truth is available. The method relies on a prior knowledge about the ratio of each structure, which may be not accurate and reliable to deal with the unseen target domain with large label distribution shift. The strong assumption of the label ratio seems to be plausible only when the label distribution has a small or moderate shift, and is insufficient to deal with more general cases.

2, The authors should consider to report the std values and p-values in the results, making the results more informative to obtain more solid conclusions.

3, Please provide more details about the model update and inference strategies.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 4xBE (1/2)

We greatly appreciate your constructive and insightful comments. We address all weaknesses below and will incorporate the corresponding changes into the revised manuscript.


W1. Supplementary Explanation of Covariate and Label Shifts.

Covariate shift originates on the acquisition side. Different centers, scanners, and acquisition protocols render the same tissue with different contrast, noise, and intensity, so the same anatomy is imaged differently. Its consequence is that the low-level statistics the model relies on no longer match the source, which scatters the predicted spatial structure and degrades boundary quality. The proportion of each structure stays unchanged.

Label shift originates on the population side. Tumor biology differs systematically between adult glioma and pediatric cohorts, so the composition of tumor sub-regions differs across populations. Its consequence is that predictions inherit the source composition and systematically under-represent the rarer foreground classes, and this bias propagates through the nested regions. The imaging appearance of each sub-region stays unchanged.


W2. The Innovation of SCA.

The central role of SCA is to supply reliable predictions for hierarchical label shift correction. Correction has to work from the model's own predictions, and those predictions are already distorted by the covariate component, so adaptation has to come before correction. SCA is that adaptation stage. It first flips the volume and requires the prediction and its mirrored copy to agree, then removes confident responses outside the tumor body. Our contribution is to assign a known regularizer this role, using it to address a problem that our setting makes unavoidable.

The flip makes the prediction confident and structurally self-consistent, so a disagreement with its mirrored copy reflects a misplaced structure rather than noise. The consistency signal brings in no external information, and what SCA recovers is where the structure sits and how far it extends, not a smoother probability map.


W3. The Innovation of HLSC.

We thank the reviewer for this assessment. Ratio-based correction is an established idea. The new knowledge in HLSC lies in the mechanism that makes correction possible in a source-free stream, where covariate shift and label shift arrive together and no target statistics are available. This mechanism, rather than a more accurate composition value, is the contribution we claim.

The two shifts are coupled and cannot be handled independently, a point that the manuscript states only briefly. Correcting label shift requires an estimate of πT\pi _T, and in a source-free stream that estimate can only come from the model's own predictions, which covariate shift has already distorted. Adapting for covariate shift alone raises the mirror problem, because the adaptation objective then treats composition-biased predictions as reliable supervision. Either ordering fails. HLSC resolves this by replacing estimation with a structural test. SCA first restores predictions that the covariate component no longer corrupts, and HLSC then checks the measured region ratios against loose one-sided bounds and applies a bounded, gradient-free logit correction only where a bound is crossed. No target prior is estimated, so the coupling has nothing left to corrupt.

Two measurements support this mechanism. The bounds need only be coarse, since sweeping them over a 7 × 7 × 7 grid changes Mean Dice by at most 1.74 points and every configuration stays within 0.90 points of the deployed one (Appendix C.2), so no dataset-specific calibration is needed. HLSC adds 7.78 in TC Dice and reduces HD95 by 15.62 mm on top of SCA, and the same bounds transfer across the three evaluation datasets without adjustment (Table 1).

The contribution is a change in what correction requires. We need neither detailed clinical assessments from experts nor a statistical analysis of a large patient cohort. A few coarse structural relations suffice, and the same relations hold across datasets. We hope this makes clear that the new knowledge is the structural test itself, which works where composition statistics cannot be estimated.

AuthorsOpenReview ↗

Response to Reviewer 4xBE (2/2)

W4&Q2. Analysis of Experimental Results.

We have added per-case paired tests with per-case improvement counts. HRAC improves over the source model by 10.68 in mean Dice (p = 1.1e-10) and 58.89 mm in HD95 (p = 1.7e-10), and over TEGDA by 2.60 in mean Dice (p = 5.5e-9), 11.30 in TC Dice (p = 9.4e-16), and 21.74 mm in HD95 (p = 4.3e-9), with 79 to 86 of 99 cases improving in each comparison.

The gain of HLSC concentrates on TC, and it is large. Adding HLSC on top of SCA gains 7.78 in TC Dice, together with 1.05 in mean Dice and 15.62 mm in HD95. TC is exactly where the population shift acts and the weakest region for the source model (32.88 Dice). The WT and ET decreases are the price of this concentration, and neither indicates a failure of the correction. The WT decrease is slight. WT always equals TC plus ED, and rTC+rED≡1r_{TC} + r_{ED} \equiv 1 holds in all cases. Reallocation between core and edema does not change total foreground volume, so the correction leaves WT nearly in place, and with the source WT already close to GT, there is little room to gain. The ET decrease is a side effect of the TC correction, not the objective of the ET term. The correction uses rET=PET/PTCr_{ET} = P_{ET} / P_{TC}, whose denominator is the inflated TC, and raising NCR compresses rETr_{ET}, which triggers the ET term and raises ET predictions. Experimentally, the resulting false positives arise almost exclusively in cases that should contain no ET.

ComparisonpImproved/Worsened
HRAC vs w/o adaptation , mean Dice1.1e-1081/18
HRAC vs w/o adaptation , mean HD951.7e-1084/15
HRAC vs TEGDA, mean Dice5.5e-979/20
HRAC vs TEGDA, mean HD954.3e-980/19
HRAC vs TEGDA, TC9.4e-1686/13
HRAC vs TEGDA WT0.02641/58
HRAC vs TEGDA ET0.26234/65

W5&Q3. Supplementary Experimental Details.

We thank the reviewer for pointing out that the test-time configuration was not fully specified. We provide it below and will make it explicit in the revised manuscript. Full training details are in Section 4.1 and Appendix B.2.

The source model is a 3D U-Net trained on the BraTS 2023 GLI training set (1,251 cases). During test-time adaptation only the BatchNorm affine parameters are updated and all other weights stay frozen. Each target volume receives 2 update steps (Adam, learning rate 3e-4, gradient clipping 1.0), followed by one gradient-free logit bias correction from HLSC (α = 4.0, ρ = (0.4, 0.6, 0.6)), and the affine parameters persist across the stream. First, inference uses whole-volume 3D forward passes, with no sliding window and no patches, so patch size does not apply and the inference batch size is 1. Second, the update iterations are 2 per volume for HRAC and Tent, 1 step for EATA, SAR, TEGDA, and SmaRT, while Undoing UDA and BLDA perform output-distribution or posterior correction with no per-step gradient updates. Third, the comparison is fair. All methods share the same source model, the same whole-volume inference protocol, and the same stream order, and a step-count sensitivity check shows that baseline mean Dice barely changes between 1 and 2 steps so the conclusion holds under either budget.

Method1 step2 steps
EATA57.5457.77 (+0.23)
SAR57.3957.19 (-0.20)
TEGDA58.8958.87 (-0.02)
SmaRT55.8155.97 (+0.16)

Q1. Applicable Scenarios of HRAC.

HRAC never estimates πTπ_T and uses no ratio value as a correction target, so the concern about mis-specified priors does not apply. The correction is driven by the measured deviation from loose structural bounds, and it strengthens as that deviation grows.

The mechanism is monotonic and bounded. As the measured ratios deviate further from the bounds, the deficits grow and the correction strengthens accordingly, while α caps the magnitude and prevents overshoot. The 343 sweep in W3 shows that the bounds need only coarse specification, with Mean Dice varying by at most 1.74 points.

The larger shift confirms this behavior. BraTS 2025 PED carries a larger shift than 2023 PED, and HRAC gains 11.56 over the source model (p = 1.2e-18) and 2.94 over TEGDA (p = 6.8e-10), with a TC gain of 25.15 (p = 5e-37). The mechanism accounts for the growth, since the source model underestimates the core more severely in this dataset and the TC gain rises from 17.5 to 25.2. In the reverse direction with PED as source and GLI as target, entropy-based methods collapse (Tent 12.34, TEGDA 3.49, SmaRT 37.80, w/o adaptation 51.10), while HRAC reaches the best mean Dice of 55.45 and the best mean HD95 of 30.11 mm (Table 5).

HRAC therefore applies to a source-free stream in which the two shifts co-occur, including a reversed direction of the compositional change. Its correction strengthens with the shift, does not depend on an accurate prior, and is not tied to a particular transfer direction.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.