acceptodds
Under review as a conference paper at ICLR 2027

Reliability-Gated Test-Time Adaptation for Cross-Domain 2D Medical Image Segmentation

Abstract

Medical image segmentation models often encounter acquisition conditions that differ from those used for training. Label-free test-time adaptation (TTA) can respond to such shifts, but entropy minimization and consistency objectives usually apply the same update pressure to every pixel. This is risky near lesion boundaries and in low-contrast regions, where an incorrect pseudo-label can reinforce the initial error. We introduce RELIABILITY-GATED TTA, a parameter-efficient TTA procedure that estimates pixel reliability from predictive entropy, alignedview disagreement, and boundary disagreement. The update is applied to reliable pixels, while an immutable source-model anchor constrains the remaining regions. Only normalization affine parameters are updated, allowing the procedure to attach to a lightweight 2D segmentation backbone. In the headline seed-7 run, Reliability-TTA lowers DDTI AURC from 0.2859 to 0.2682 and improves AURC under brightness and speckle perturbations, while Gaussian blur reverses the trend. Across the three seeds, overlap accuracy is largely unchanged on DDTI and can decrease on STU, with seed 8 changing from 0.8807 to 0.8009. These results show that spatial reliability cues can modify selective risk under some appearance shifts, but they do not provide uniform accuracy recovery. The evidence characterizes a shift-dependent operating regime for label-free 2D TTA with a normalization-affine update subspace.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Test-Time Adaptation, Gated adaptation, Cross-domain, Medical image segmentation

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_BvWfOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates Reliability-Gated Test-Time Adaptation (Reliability-TTA) for cross-domain 2D medical image segmentation. The proposed framework constructs a pixel-wise reliability gate by combining predictive entropy, disagreement between spatially aligned augmented views, and boundary-response disagreement. Reliable pixels contribute to entropy minimization and consistency learning, their local neighborhoods support boundary consistency, and the complementary regions are regularized using source-posterior anchoring. Adaptation is restricted to normalization-affine parameters. Experiments are conducted using MSCANet trained on TN3K and evaluated on two target datasets, DDTI and STU, across three independently trained checkpoints. The proposed adaptation consistently decreases Dice on DDTI and produces only marginal improvements on STU.

Critical strengths
  1. Clear problem definition
  2. Careful experimental protocol
  3. Informative analysis of spatial reliability and adaptation effectiveness
Critical weaknesses
  1. Figure 1 contains confusing connections and malformed or uninterpretable mathematical symbols, making it difficult to understand how the components interact.
  2. The proposed adaptation consistently decreases Dice on DDTI and produces only marginal improvements on STU.
  3. Insufficient experimental results and limited validation of the claimed reliability mechanism.
  4. Lack of sufficient research on TTA and absence of related works.
  5. The writing is poor.
Questions for authors
  1. Please demonstrate the effectiveness of the proposed method.
  2. Please conduct more experiments, including more ablation studies, comparisons, and discussions.
  3. Please conduct thorough research in the field of TTA.
Reviewer_NUKbOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a reliability-gated test-time adaptation method for medical image segmentation. It combines prediction entropy, consistency across augmented views, and boundary response differences to estimate pixel-level reliability, using reliable regions to guide model updates while keeping the source model as an anchor. Experiments on thyroid ultrasound datasets show that the spatial gating identifies some error-prone regions, but yields little to no improvement in segmentation performance.

Critical strengths

The paper clearly reports its experimental results and limitations, including the lack of consistent segmentation improvements.

Critical weaknesses
  1. Several figures are confusing and poorly organized, making the method and experimental results difficult to interpret. This raises concerns about the clarity and overall presentation of the work.
  2. The error enrichment results lack comparisons with simple baselines, such as random pixel selection with the same coverage. It is therefore unclear whether the proposed reliability gating provides meaningful benefits beyond basic spatial selection.
Questions for authors

Can the authors clarify the design of the figures and improve their readability, particularly where the method and experimental results are difficult to interpret?

Reviewer_y84bOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Reliability-Gated TTA, a test-time adaptation method for 2D medical image segmentation. It builds a per-pixel reliability gate from three cues: predictive entropy, disagreement between intensity-augmented views, and boundary disagreement. Gated pixels drive entropy and consistency losses, while the remaining pixels are anchored to the source prediction; only normalization-affine parameters are updated. The source model is a thyroid nodule segmenter trained on TN3K and evaluated on DDTI and a 42-image collection labeled STU. Across three source checkpoints, the method slightly decreases DDTI Dice and yields negligible STU gains. The authors therefore present the work mainly as a diagnostic finding: the gate concentrates errors in its excluded region, but the update does not convert this into better segmentation.

Critical strengths

1. Transparent reporting of a negative result.

  • The paper reports seed-wise paired changes for three independently trained checkpoints (Table 2).
  • It keeps case-level bootstrap intervals distinct from checkpoint variability.
  • It states that target masks never enter the adaptation computation.
  • It acknowledges that the method decreases DDTI Dice for every seed.
Critical weaknesses

1. The method provides no segmentation benefit, and the diagnostic conclusion that replaces it is not established.

  • The method does not help on this task, and a simpler baseline does better: On DDTI the method changes Dice by −0.00159 ± 0.00089, negative for all three seeds; on STU by +0.00025 ± 0.00019 (Table 2). AdaBN has higher mean Dice on both datasets. These effects are one to two orders of magnitude smaller than the variation across source checkpoints (DDTI source Dice ranges from 0.698 to 0.765).
  • The update appears too small to test the objective: The update consists of two Adam steps on normalization-affine parameters at a learning rate that is not reported. In the seed-9 sensitivity analysis (Table 3), every case-wise variant stays within −0.00011 to +0.00122 Dice of the default on DDTI. This includes:
    • disabling gating entirely (keep ratio 100%: −0.00002);
    • removing the consistency or boundary loss (~0). By contrast, the continual stream, which accumulates updates, changes Dice by +0.00536. This pattern indicates that the case-wise update barely improves the model. The paper's diagnosis, a "selector–update failure mode," therefore cannot be distinguished from insufficient adaptation. Grid sweeps over learning rate and step count would be necessary.
  • The spatial diagnostic is largely expected: With a 50% gate, the maximum error enrichment is 2, as the paper notes. Excluding the highest-uncertainty half of pixels will naturally capture most errors, because segmentation errors concentrate near predicted boundaries where entropy is high. The spatial diagnostic does not compare the three-cue score against simpler gates (entropy only, random, or distance to the predicted boundary). Boundary-band enrichment near 1 (1.008 on DDTI, 1.002 on STU) suggests the composite score adds no separation beyond this.

2. The evaluation is narrow, and the validity of key data choices is doubtful.

  • Narrow scope: The study uses one lightweight backbone, MSCANet (originally a colorectal tumor segmenter), one source dataset, and two targets. The mechanism, anchor, sensitivity, corruption, and readout analyses all use a single checkpoint. Beyond AdaBN and entropy minimization, no established TTA methods are compared, including the single-image medical methods the paper itself identifies as the closest protocol comparison (Sec. 2.2). An existing large-scale medical TTA benchmark already includes the TN3K->DDTI setting and could provide this comparison.
  • The STU target may not be a thyroid dataset: STU is described only as a "locally held collection," included "only if its foreground masks denote the same anatomical target as TN3K" (Sec. 4.1). The paper states that every target collection is audited before inclusion, but it neither describes the collection nor reports the audit outcome. A public dataset named STU consists of 42 breast ultrasound images from Shantou University. If this is the dataset used, it contains breast lesions, not thyroid nodules, violating the paper's own inclusion criterion (cf. Sec. 6 on gland masks).
  • The corruption study may include training images: It is described as using "3,493 TN3K source-validation images" (Sec. 4.6, Table 4). TN3K contains 3,493 images in total, with an official split of 2,879 training/validation images and 614 test images. This evaluation set therefore appears to include the source training images.
  • Reported numbers conflict: The anchor ablation reports DDTI Dice changing from 0.74080 to 0.74089 when the anchor is removed (Sec. 4.5). However, Table 3 and Table 2 imply that the default configuration, which includes the anchor, scores 0.74089.

The overview figure is also messy, and over-detailed with equations that are also rendered incorrectly and font size too small to read.

Questions for authors
  1. Update magnitude: What learning rate is used? What is the norm of the parameter change after two steps? Do entropy TTA and Reliability-Gated TTA diverge in outcome under larger learning rates or more steps? Without this, the failure cannot be attributed to the objective rather than to under-adaptation.
  2. Gate baselines: What error capture and enrichment do an entropy-only gate, a random 50% gate, and a gate based on distance to the predicted boundary achieve? This determines whether the three-cue score contributes anything to the diagnostic finding.
  3. STU identity: What is the STU collection: its source, acquisition device, number of patients, and the anatomical structure its masks denote? Is it the 42-image Shantou University breast ultrasound dataset? If so, why is it a valid target domain for a thyroid nodule segmenter?
  4. TN3K split: How does the 3,493-image "source-validation" set relate to TN3K's official training and test splits? Were any images used to train or select the checkpoints included in the corruption study?
  5. Anchor ablation: Please reconcile the with-anchor DDTI Dice of 0.74080 (Sec. 4.5) with the default 0.74089 in Table 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.