acceptodds
Under review as a conference paper at ICLR 2027

Explicit Adaptation Regulation for Data-Efficient Medical Image Segmentation

Abstract

Data-efficient medical image segmentation requires accurate in-domain predictions and reliable transfer to unseen datasets from limited source-domain annotations. Existing approaches largely regulate adaptation through latent features, leaving the connection between the current pixel–text assessment and subsequent updates to be learned indirectly through task supervision. With supervision concentrated on the final mask, this relationship can remain weakly constrained and become dependent on source-specific patterns, limiting performance under scarce supervision and domain shift. To bridge this gap, we introduce Explicit Adaptation Regulated Segmentation (ExpARSeg), which makes local pixel-text states explicit regulators of representation adaptation and spatial prediction. Specifically, ExpARSeg employs Evidential Pixel-Text (EPT) Adapters to condition bidirectional updates within frozen vision and text encoders on local support and commitment. It further integrates Selective Spatial Rendering (SSR) to direct text-conditioned visual corrections toward low-commitment locations identified by the current margin, with initial residual responses guiding subsequent point selection. Experiments across four imaging modalities and ten held-out target datasets show that ExpARSeg improves sample efficiency and domain generalization over state-of-the-art baselines using implicit adaptation regulation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
medical image segmentation, vision-language models, explicit adaptation regulation, data-efficient learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_UYdkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents ExpARSeg for medical image segmentation with limited source annotations and transfer to unseen datasets. It uses local pixel–text compatibility to control bidirectional feature updates in frozen vision and text encoders. A two-stage refinement module selects locations with low prediction margins and applies text-conditioned corrections using local image details. The model is trained with segmentation losses and a target-conditioned KL regularizer. Experiments span four imaging modalities and ten target datasets. ExpARSeg achieves the highest DSC in 11 of 12 source-data settings and the highest average target DSC in three of four modality groups.

Critical strengths
  1. The EPT adapter provides a concrete way to make cross-modal updates depend on local pixel–text compatibility. It uses foreground support and commitment to route the exchanged features, giving adaptation an explicit task-related bias. This is a useful design contribution for learning from limited mask supervision.
  2. The evaluation spans four imaging modalities and ten held-out target datasets, with a shared pretrained backbone for CLIP-based methods. Table 1 lists improvements across all four modalities at the 10% annotation budget. This provides useful evidence that the approach can benefit segmentation with limited annotations across several tasks.
Critical weaknesses
  1. Appendix A.1 states that prompt construction uses attributes from images, masks, and labels. All text-based methods share these prompts, so this does not invalidate their comparison. My concern is whether ExpARSeg still performs better when such descriptions are unavailable. The generic-prompt experiment helps, but it reports recall at a controlled background false-positive rate rather than DSC. It does not show whether the reported segmentation gains remain with generic prompts.
  2. Table 2 lists transfer results only at the 100% source annotation budget. It remains unclear whether ExpARSeg also outperforms the main baselines on target datasets when trained with 10% or 50% of the source annotations. Figure 5 provides useful feature analyses at these budgets, but does not answer this segmentation question.
Questions for authors
  1. With generic task prompts replacing the supplied descriptions at test time, what are the target DSC scores for ExpARSeg and MedCLIPSeg under the same evaluation protocol as Table 2? The recall analysis in Figure 11 does not show whether the segmentation advantage remains.
  2. Does the mean DSC in Figure 6 include target datasets? Please clarify which datasets it averages over. Are target DSC comparisons with the main baselines available for models trained with 10% or 50% of the source annotations?
  3. How well does SSR recover lesions that receive confidently negative coarse margins on target datasets? Its point selection favors small absolute margins, which may leave these missed regions unselected.
6 comments
AuthorsOpenReview ↗

Response to Reviewer UYdk [Q1]: Segmentation Performance using Generic Task Prompts

We thank the reviewer for the precise and well-argued review, and we hope our responses can deal with the main concerns.

Q1. Whether the reported segmentation gains remain with generic prompts?

The original benchmark in our main submission allows us to compare adaptation methods under a common level of textual assistance. Now we extend the generic-prompt evaluation protocol in Figure 11 to further assess final segmentation quality using DSC. Specifically, we predefine six templates using only the imaging modality and segmentation target for each modality, such as “Segment the polyp regions in this endoscopic image.” All methods share the same pool across annotation budgets, seeds, and source and target datasets. Template construction, assignment, and weighting are independent of patient annotations and validation results. At deployment, the task and modality determine the pool, so the same instructions can be reused for every patient. We have also included results from two extra random seeds for a more comprehensive evaluation. Comparisons between our method and the strongest baseline MedCLIPSeg are as follows:

10% Data:

MethodMedCLIPSegExpARSeg (ours)
Overall68.86±0.9470.84±0.53
BUSI57.25±4.3259.63±3.33
BTMRI78.17±0.8880.35±0.07
Kvasir-SEG77.43±1.1478.62±1.36
ISIC90.18±0.2590.23±0.10
BUSUC73.54±3.7376.28±0.99
BUSBRA52.42±6.5756.36±1.98
BUID64.48±2.8169.32±1.04
UDIAT64.43±3.4667.18±0.55
CVC-ColonDB53.23±1.1855.92±1.09
CVC-ClinicDB66.11±0.6467.45±0.72
CVC-30071.10±3.2274.17±1.63
BKAI62.73±2.0864.83±1.61
BRISC70.82±0.5373.46±0.62
UWaterlooSkinCancer82.17±2.2677.87±2.48

50% Data:

MethodMedCLIPSegExpARSeg (ours)
Overall75.93±0.1677.16±0.31
BUSI67.96±2.9272.61±2.20
BTMRI84.08±0.1985.36±0.26
Kvasir-SEG84.71±0.7586.09±0.36
ISIC91.47±0.1591.76±0.27
BUSUC81.01±1.2682.85±1.17
BUSBRA69.02±1.0469.07±2.29
BUID75.51±0.9476.90±0.75
UDIAT71.38±1.4071.44±3.90
CVC-ColonDB62.12±0.7564.15±0.24
CVC-ClinicDB72.73±1.4773.36±1.22
CVC-30077.90±1.3680.61±0.76
BKAI68.42±0.6869.24±0.43
BRISC72.48±0.2174.45±1.23
UWaterlooSkinCancer84.30±1.3382.36±3.75

100% Data:

MethodMedCLIPSegExpARSeg (ours)
Overall78.53±0.1579.44±0.20
BUSI74.62±2.4078.35±2.00
BTMRI86.06±0.5987.92±0.31
Kvasir-SEG87.44±0.2587.55±0.44
ISIC91.53±0.2291.84±0.06
BUSUC84.71±0.3484.92±0.47
BUSBRA73.15±0.8973.60±1.54
BUID77.31±0.6077.43±1.97
UDIAT74.49±1.6576.07±0.99
CVC-ColonDB66.41±0.3867.24±0.10
CVC-ClinicDB76.59±2.1776.75±1.18
CVC-30080.63±1.5882.03±1.27
BKAI72.24±0.5771.86±0.36
BRISC72.02±2.3175.03±0.55
UWaterlooSkinCancer82.26±0.8281.54±1.08

Averaged over the ten target datasets, ExpARSeg obtains DSC scores of 68.29±0.38, 74.44±0.48, and 76.65±0.46 at the 10%, 50%, and 100% annotation budgets, respectively, while MedCLIPSeg obtains 66.10±1.52, 73.49±0.22, and 75.98±0.32. ExpARSeg has the higher three-seed mean on 9/10, 9/10, and 8/10 target datasets, with the largest average gain at 10% annotations. Examples at this budget include BUID, where DSC rises from 64.48 to 69.32, CVC-300, from 71.10 to 74.17, and BRISC, from 70.82 to 73.46. We also report performance on the four source datasets, where ExpARSeg continues to lead across the board. In summary, the results of this supplementary experiment show that ExpARSeg achieves higher DSC than MedCLIPSeg on most target datasets when both methods are trained and tested with generic task prompts. Its advantage therefore extends to a practical setting where prompts specify the segmentation task and require no patient-specific annotations.

AuthorsOpenReview ↗

Response to Reviewer UYdk [Q2]: Segmentation Performance under Different Annotation Budgets

Q2. Provide comparisons for models trained with 10% or 50% of the source annotations.

Figure 6 reports the mean DSC on the source test sets of BUSI, Kvasir, BTMRI, and ISIC, separately for each modality using 10% of the source training annotations. Mean DSC refers to that each value is averaged over the test images within the corresponding dataset. Target datasets are not included.

Here we provide DSC comparisons with the strongest baseline MedCLIPSeg under different annotation budgets. Results are as follows:

10% Data:

MethodMedCLIPSegExpARSeg
Overall70.42±0.7173.07±0.66
BUSI57.73±2.1569.76±1.26
BTMRI83.24±0.1682.51±1.90
Kvasir-SEG75.37±2.4180.67±1.30
ISIC90.45±0.1790.56±0.29
BUSUC77.68±1.7776.09±3.85
BUSBRA58.03±2.9358.99±6.14
BUID69.98±2.3566.26±3.15
UDIAT67.13±0.7370.19±2.52
CVC-ColonDB53.17±2.3358.46±2.61
CVC-ClinicDB64.19±0.8868.89±1.55
CVC-30073.09±4.7876.61±0.86
BKAI63.12±0.5665.49±1.98
BRISC68.51±9.2275.74±1.52
UWaterlooSkinCancer84.18±0.4282.80±2.11

50% Data:

MethodMedCLIPSegExpARSeg
Overall78.89±1.2380.28±0.35
BUSI80.37±0.6680.49±0.60
BTMRI87.01±0.1987.90±0.24
Kvasir-SEG87.97±0.5487.97±0.33
ISIC91.76±0.4292.19±0.27
BUSUC81.54±1.3582.01±2.16
BUSBRA66.38±6.6474.46±1.79
BUID75.74±3.4379.88±1.74
UDIAT72.32±3.8473.57±3.38
CVC-ColonDB68.43±0.6168.10±0.64
CVC-ClinicDB76.67±1.2576.53±1.18
CVC-30079.80±0.7980.64±1.36
BKAI74.42±1.3774.65±2.56
BRISC77.69±1.7180.50±1.52
UWaterlooSkinCancer84.34±0.6185.09±2.43

100% Data:

MethodMedCLIPSegExpARSeg
Overall80.11±1.3281.98±0.42
BUSI84.09±0.9285.08±1.10
BTMRI88.69±0.0488.98±0.27
Kvasir-SEG89.95±0.5989.81±0.15
ISIC92.33±0.0292.59±0.04
BUSUC79.71±2.6083.60±1.88
BUSBRA67.04±3.3174.87±1.18
BUID76.80±4.7180.12±1.49
UDIAT72.88±2.6875.29±4.11
CVC-ColonDB72.83±0.5873.17±0.59
CVC-ClinicDB80.21±0.4280.85±0.15
CVC-30082.96±0.5683.39±0.67
BKAI77.89±1.4979.48±1.20
BRISC77.75±4.6278.48±4.74
UWaterlooSkinCancer78.49±4.2482.03±2.41

With 10% source annotations, ExpARSeg averages 69.95 DSC across the ten target datasets, compared with 67.91 for MedCLIPSeg. This is a gain of 2.04 percentage points (pp), with ExpARSeg ahead on 7 of 10 targets. All four endoscopy targets improve: CVC-ColonDB by 5.29 pp, CVC-ClinicDB by 4.70 pp, CVC-300 by 3.52 pp, and BKAI by 2.37 pp. The largest gain is on BRISC (+7.23 pp), while UDIAT and BUSBRA improve by 3.06 pp and 0.96 pp. The gains therefore extend across several target datasets and modalities at this extremely limited annotation budget.

With 50% source annotations, the target mean DSC is 77.54 for ExpARSeg and 75.73 for MedCLIPSeg, giving an advantage of 1.81 pp. The largest improvements are on BUSBRA (+8.08 pp) and BUID (+4.14 pp), followed by BRISC (+2.81 pp). Smaller gains occur on UDIAT, BUSUC, CVC-300, BKAI, and UWaterlooSkinCancer. The two datasets where MedCLIPSeg remains ahead are CVC-ColonDB and CVC-ClinicDB, with differences of just 0.33 pp and 0.14 pp. Thus, the higher target average is accompanied by gains on most individual datasets.

With 100% source annotations, ExpARSeg reaches 79.13 mean target DSC versus 76.66 for MedCLIPSeg, with an improvement of 2.47 pp. It leads on all ten target datasets, with gains ranging from 0.34 to 7.83 pp. These include 7.83 pp on BUSBRA, 3.89 pp on BUSUC, 3.54 pp on UWaterlooSkinCancer, and 3.32 pp on BUID.

To sum up, ExpARSeg achieves higher average target DSC at all three annotation budgets and leads on 25 of the 30 target comparisons settings. Its advantage extends from scarce source annotations to full supervision, with gains on most targets at 10% and 50% and on all targets at 100%.

AuthorsOpenReview ↗

Response to Reviewer UYdk [Q3]: Effect of SSR on Confidently Negative Regions

Q3. Whether the point selection of SSR only favors small absolute margins?

SSR can update a missed region through corrections applied at neighboring locations. Half-resolution corrections are interpolated onto the full-resolution grid, allowing their effects to reach pixels beyond the directly selected points. The updated margins then guide the next round of point selection. An initially confident false negative within this spatial support can therefore receive a correction and become a candidate for further refinement. At both stages, fine visual details and text conditioning provide information for revising the coarse prediction.

The BRISC and UWaterloo target cases in Figure 8 show how this refinement improves lesion coverage. Parts of the annotated lesions are absent from the coarse masks but recovered in the final predictions, with the cyan regions showing where SSR changes the predicted labels. The final masks extend into the missed foreground and follow the GT contours more closely. These examples support SSR’s ability to recover missed lesion regions through its two-stage spatial refinement.

Reviewer_RtWfOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ExpARSeg, a framework for data-efficient and generalizable medical image segmentation. The method introduces Evidential Pixel-Text (EPT) Adapters to regulate bidirectional vision-language interactions based on local pixel-text states, and a Selective Spatial Rendering (SSR) module to refine segmentation predictions at uncertain locations. Experiments across four modalities and different annotation budgets demonstrate improved segmentation performance. Additional ablation studies are provided to investigate the proposed mechanisms.

Critical strengths

The paper addresses data efficiency and domain generalization in medical image segmentation under limited annotations and distribution shifts, which are important problems. The paper provides a relatively comprehensive evaluation across multiple imaging modalities, annotation budgets, and target datasets. The proposed method achieves competitive results. The authors also provide extensive ablation studies and additional analyses.

Critical weaknesses

1. Limited novelty in method and insufficient evidence for the claimed mechanism

The proposed framework seems to be a combination of existing techniques rather than a new methodology. MedCLIPSeg has already introduced uncertainty-aware bidirectional vision-language adaptation, and PointRend has explored uncertainty-guided point selection and coarse-to-fine segmentation refinement. The distinction of the proposed Explicit Adaptation Regulation from existing approaches is not sufficiently convincing. Particularly, the evidential state is deterministically obtained from the pixel-text similarity, and whether the improvement comes from evidential modeling or simply from confidence-based feature regulation is unclear. Additionally, the experiments haven't fully supported that the claimed explicit regulation mechanism really improve the results. Some key ablations change multiple components at the same time, and the feature analyses mainly demonstrate correlations between the proposed architecture and improved representations. There is no a directly comparison to verify the isolated explicit regulation.

2. Concerns about experimental validity and the consistency of reported improvements

The experimental protocol has several issues that may affect the validity of the conclusions. For example, Appendix A.1 states that text prompts are from the MedCLIPSeg benchmark. But MedCLIPSeg's original supplementary material describes a prompt-generation procedure that uses ground-truth masks to extract lesion attributes. The authors need to clarify whether such ground-truth information is used for target-domain evaluation. In addition, the reported improvements are not consistent across all modalities, and some baseline results show large variations across training-data ratios. The authors should use multiple random seeds for evaluations to see the statistical significance of the relatively small improvements.

In overall, though the empirical results are promising, I'm not fully convinced that the paper has well established a sufficiently strong methodological contribution other than combining existing techniques. I currently lean toward a weak rejection, but my assessment may change depending on the authors' responses to the questions, especially those for the evaluation protocols and the claimed main contributions.

Questions for authors

1 . Were the source and target test prompts generated using ground-truth masks or mask-derived attributes? Please report the DSC of ExpARSeg and MedCLIPSeg using ground-truth-independent prompts.

2 . The authors should justify that the Dirichlet/Beta formulation is necessary and is not over-engineered. Could the authors replace it with a much simpler deterministic confidence-based mechanism, such as a sigmoid gating, that has the same regulation functionality, while keeping the remaining network architecture unchanged? A direct performance comparison would help us better understand whether the additional module complexity is really needed.

3 . Table 1 shows non-monotonic baseline performance. For example, CausalCLIPSeg on Brain MRI drops from 76.32 to 64.44 DSC, and CAT-Seg on Breast Ultrasound drops from 70.09 to 59.57, when supervision increases from 10% to 50%. Could the authors clarify these unexpected results?

4 . MedCLIPSeg achieves only 55.51 DSC on BUSI with 10% supervision in Table 1. But the original MedCLIPSeg paper reported 68.66. Could the authors explain the gap?

5 . To assess the statistical reliability without many additional experiments, I suggest the authors reporting mean and standard deviation over at least three training seeds for two most representative comparisons: (1) ExpARSeg vs. CAT-Seg on BUSI with 10% supervision, and (2) ExpARSeg vs. MedCLIPSeg on BUSI-to-target Breast Ultrasound generalization. More reported results with mean and variance are welcomed.

6 . Could the authors report the DSC separately for all ten target datasets and explain why the improvements over MedCLIPSeg are concentrated in Breast Ultrasound and Skin Dermatoscopy, while being marginal or negative in the other two modalities? These results may require narrowing down the claim of domain generalization.

7 . Table 3(i) shows that the TC-KL regularizer contributes to a substantial part of the overall improvement over MedCLIPSeg. Could the authors clarify how this affects the claimed contribution of the proposed explicit regulation mechanism?

8 comments
AuthorsOpenReview ↗

Response to Reviewer RtWf [Q1]: Segmentation Performance using Mask-independent Prompts

We thank the reviewer for a detailed and constructive review, and we hope our responses can solve the main concerns.

Q1. Please report the DSC of ExpARSeg and MedCLIPSeg using ground-truth-independent prompts.

Our main experiments follow the released MedCLIPSeg benchmark, where the source and target test prompts were generated using mask-derived attributes. This shared setting allows us to compare adaptation methods under a common level of textual assistance.

To examine performance under a more practical setting, we have evaluated performance using instructions specified solely from the segmentation task. Specifically, we write six templates using only the modality and target category for each modality, such as “Segment the polyp regions in this endoscopic image.” All methods are retrained with these templates, cycling through them during training and averaging their predictions with equal weights at inference. The templates and weights are fixed across annotation budgets, seeds, and source and target datasets. Their construction and use are independent of patient annotations and validation results. In practice, the requested task and modality determine the prompt pool, allowing the same instructions to be reused for every patient. We compare our ExpARSeg with MedCLIPSeg as follows:

10% Data:

MethodMedCLIPSegExpARSeg (ours)
Overall68.86±0.9470.84±0.53
BUSI57.25±4.3259.63±3.33
BTMRI78.17±0.8880.35±0.07
Kvasir-SEG77.43±1.1478.62±1.36
ISIC90.18±0.2590.23±0.10
BUSUC73.54±3.7376.28±0.99
BUSBRA52.42±6.5756.36±1.98
BUID64.48±2.8169.32±1.04
UDIAT64.43±3.4667.18±0.55
CVC-ColonDB53.23±1.1855.92±1.09
CVC-ClinicDB66.11±0.6467.45±0.72
CVC-30071.10±3.2274.17±1.63
BKAI62.73±2.0864.83±1.61
BRISC70.82±0.5373.46±0.62
UWaterlooSkinCancer82.17±2.2677.87±2.48

50% Data:

MethodMedCLIPSegExpARSeg (ours)
Overall75.93±0.1677.16±0.31
BUSI67.96±2.9272.61±2.20
BTMRI84.08±0.1985.36±0.26
Kvasir-SEG84.71±0.7586.09±0.36
ISIC91.47±0.1591.76±0.27
BUSUC81.01±1.2682.85±1.17
BUSBRA69.02±1.0469.07±2.29
BUID75.51±0.9476.90±0.75
UDIAT71.38±1.4071.44±3.90
CVC-ColonDB62.12±0.7564.15±0.24
CVC-ClinicDB72.73±1.4773.36±1.22
CVC-30077.90±1.3680.61±0.76
BKAI68.42±0.6869.24±0.43
BRISC72.48±0.2174.45±1.23
UWaterlooSkinCancer84.30±1.3382.36±3.75

100% Data:

MethodMedCLIPSegExpARSeg (ours)
Overall78.53±0.1579.44±0.20
BUSI74.62±2.4078.35±2.00
BTMRI86.06±0.5987.92±0.31
Kvasir-SEG87.44±0.2587.55±0.44
ISIC91.53±0.2291.84±0.06
BUSUC84.71±0.3484.92±0.47
BUSBRA73.15±0.8973.60±1.54
BUID77.31±0.6077.43±1.97
UDIAT74.49±1.6576.07±0.99
CVC-ColonDB66.41±0.3867.24±0.10
CVC-ClinicDB76.59±2.1776.75±1.18
CVC-30080.63±1.5882.03±1.27
BKAI72.24±0.5771.86±0.36
BRISC72.02±2.3175.03±0.55
UWaterlooSkinCancer82.26±0.8281.54±1.08

With ground-truth-independent prompts, ExpARSeg achieves higher mean DSC than MedCLIPSeg on 13/14, 13/14, and 12/14 datasets at the 10%, 50%, and 100% annotation budgets, respectively. Overall DSC reaches 70.84, 77.16, and 79.44, compared with 68.86, 75.93, and 78.53 for MedCLIPSeg. Beyond leading on all four source datasets at every budget, ExpARSeg also outperforms MedCLIPSeg on 9/10, 9/10, and 8/10 target datasets, directly addressing the concern about target-domain performance when annotation-derived descriptions are unavailable. Its average target DSC is 68.29, 74.44, and 76.65, compared with 66.10, 73.49, and 75.98 for MedCLIPSeg. The largest average gain occurs with 10% annotations, with improvements across several transfer settings, including BUID (64.48 to 69.32), CVC-300 (71.10 to 74.17), and BRISC (70.82 to 73.46).

To sum up, these gains above are obtained with prompts that specify only the modality and segmentation task. Such instructions can be prepared once and reused across patients, independently of their annotations. The higher DSC on most target datasets therefore shows that ExpARSeg’s advantage over MedCLIPSeg remains when training and inference use task descriptions readily available in practice.

AuthorsOpenReview ↗

Response to Reviewer RtWf [Q2]: Advantages of Dirichlet-based Formulation

Q2. Please provide direct comparisons between Dirichlet/Beta formulation and simpler mechanisms.

We have evaluated three representative mechanisms: similarity-based, confidence-based and entropy-based regulation under identical settings. Direct performance comparisons are as follows:

10% Data:

DatasetSimilarity gatingConfidence attentionEntropy uncertaintyExpARSeg
Overall69.32 ± 2.6670.48 ± 0.9270.94 ± 1.1373.07 ± 0.66
BUSI73.84 ± 1.0772.11 ± 1.6971.51 ± 3.4969.76 ± 1.26
BUSBRA55.27 ± 8.9762.04 ± 0.3461.85 ± 4.1058.99 ± 6.14
BUSUC65.16 ± 10.5671.09 ± 1.7171.82 ± 3.6276.09 ± 3.85
BUID43.26 ± 11.6655.16 ± 4.7556.69 ± 8.2166.26 ± 3.15
UDIAT63.30 ± 3.4059.94 ± 6.6163.53 ± 3.1270.19 ± 2.52
BTMRI81.77 ± 0.1482.00 ± 0.4682.18 ± 0.5282.51 ± 1.90
BRISC73.90 ± 0.5973.91 ± 0.3074.54 ± 0.4675.74 ± 1.52
Kvasir-SEG80.44 ± 0.4479.79 ± 2.1380.38 ± 0.3580.67 ± 1.30
BKAI64.20 ± 1.1563.94 ± 3.5164.00 ± 1.4465.49 ± 1.98
CVC-ColonDB57.61 ± 1.1957.03 ± 1.1156.49 ± 0.5758.46 ± 2.61
CVC-ClinicDB67.37 ± 0.6467.09 ± 3.1766.74 ± 0.8368.89 ± 1.55
CVC-30074.82 ± 1.1474.85 ± 1.4274.73 ± 1.9876.61 ± 0.86
ISIC89.65 ± 0.4989.27 ± 0.4489.30 ± 0.2790.56 ± 0.29
UWaterlooSkinCancer79.86 ± 3.9878.52 ± 1.4479.46 ± 1.7682.80 ± 2.11

50% Data:

DatasetSimilarity gatingConfidence attentionEntropy uncertaintyExpARSeg
Overall77.11 ± 1.5876.82 ± 0.4676.58 ± 0.6180.28 ± 0.35
BUSI80.43 ± 2.1780.62 ± 0.4481.32 ± 0.7580.49 ± 0.60
BUSBRA69.36 ± 2.7966.83 ± 1.9467.41 ± 1.6874.46 ± 1.79
BUSUC76.84 ± 5.8576.58 ± 1.1477.66 ± 1.2682.01 ± 2.16
BUID68.36 ± 9.7269.72 ± 5.3970.28 ± 3.6079.88 ± 1.74
UDIAT67.23 ± 4.2163.54 ± 6.3564.74 ± 8.1473.57 ± 3.38
BTMRI86.38 ± 0.2085.99 ± 0.0585.94 ± 0.4687.90 ± 0.24
BRISC78.77 ± 0.7178.10 ± 0.4078.46 ± 0.4080.50 ± 1.52
Kvasir-SEG87.42 ± 0.7487.70 ± 0.1786.90 ± 0.8587.97 ± 0.33
BKAI75.38 ± 1.0774.14 ± 1.4073.24 ± 0.6774.65 ± 2.56
CVC-ColonDB67.26 ± 1.2666.62 ± 2.2565.06 ± 0.9368.10 ± 0.64
CVC-ClinicDB75.14 ± 2.6874.42 ± 2.1173.69 ± 0.9276.53 ± 1.18
CVC-30079.76 ± 1.6581.05 ± 2.0678.92 ± 1.3980.64 ± 1.36
ISIC91.64 ± 0.3491.25 ± 0.4391.60 ± 0.1592.19 ± 0.27
UWaterlooSkinCancer75.61 ± 3.0278.91 ± 0.9776.95 ± 4.9885.09 ± 2.43

100% Data:

DatasetSimilarity gatingConfidence attentionEntropy uncertaintyExpARSeg
Overall80.11 ± 0.4977.78 ± 0.6874.79 ± 4.5681.98 ± 0.42
BUSI84.40 ± 0.1683.58 ± 0.3884.71 ± 0.3085.08 ± 1.10
BUSBRA72.71 ± 1.3866.02 ± 4.5958.32 ± 13.5174.87 ± 1.18
BUSUC81.55 ± 1.2677.67 ± 3.2861.53 ± 17.8583.60 ± 1.88
BUID76.08 ± 0.7671.39 ± 1.5753.29 ± 18.8480.12 ± 1.49
UDIAT69.12 ± 0.9862.70 ± 4.5663.62 ± 7.3475.29 ± 4.11
BTMRI87.94 ± 0.1187.48 ± 0.7487.85 ± 0.2188.98 ± 0.27
BRISC80.50 ± 0.1478.95 ± 1.3480.47 ± 0.4678.48 ± 4.74
Kvasir-SEG89.41 ± 0.2688.83 ± 0.6788.85 ± 0.9689.81 ± 0.15
BKAI78.20 ± 0.5977.65 ± 1.1177.27 ± 1.2679.48 ± 1.20
CVC-ColonDB71.12 ± 1.1670.72 ± 1.1969.85 ± 0.7373.17 ± 0.59
CVC-ClinicDB78.53 ± 0.0279.15 ± 2.5576.58 ± 0.4180.85 ± 0.15
CVC-30082.23 ± 1.2981.74 ± 1.5981.56 ± 1.0783.39 ± 0.67
ISIC92.20 ± 0.1691.97 ± 0.1991.85 ± 0.0792.59 ± 0.04
UWaterlooSkinCancer77.50 ± 5.9871.08 ± 8.7671.37 ± 14.6982.03 ± 2.41

As shown above, ExpARSeg outperforms all three simpler alternatives in overall DSC at every annotation budget, exceeding the strongest alternative by 2.13, 3.17, and 1.87 pp. For datasets-wise comparisons, it also leads on 12/14, 11/14, and 13/14 datasets, with lower variability in overall DSC across different seeds. These gains justify our explicit regulation design through measurable segmentation benefits. The evaluated simpler mechanisms offer lower overall performance at all three budgets, supporting the practical value of our proposed evidential design.

AuthorsOpenReview ↗

Response to Reviewer RtWf [Q3,Q4,Q5]: More Discussion about Experimental Validity

Q3 & Q4. Gap from the original MedCLIPSeg paper

We have carefully reviewed the experimental settings in our main submission and compared them with those in the original MedCLIPSeg paper. The only difference lies in the protocol used to select the checkpoints. Specifically, the original MedCLIPSeg paper specifies final-epoch evaluation for MedCLIPSeg, while some other baselines use the validation sets to select their final checkpoints. Instead, we have all methods select their final checkpoint using the same validation set, ensuring that comparisons are made under a more consistent protocol. We chose validation-based selection because it can retain the procedure already used by certain baselines such as LViT and gives every method the same opportunity for checkpoint selection. Given that models may easy to overfit under low data rate, validation DSC offers a common criterion based on segmentation quality on held-out source images, allowing each method to select an appropriate epoch within the same training budget. A fixed final epoch specifies training duration, whereas validation selection also considers how well the learned model performs on unseen cases.

Since we are not dealing with extreme scenarios such as zero-shot or one-shot, validation data can always be obtained by reserving a small portion of the available labeled source cohort, which is feasible in practical development and deployment scenarios. For our ExpARSeg, checkpoints selected by source validation DSC also achieve strong performance on held-out test data. On BUSI with 10% supervision, selection using only seven validation images yields a test DSC of 69.76±1.26 across three training seeds. The same selection criterion also identifies models that generalize to unseen target datasets. ExpARSeg exceeds MedCLIPSeg’s mean target DSC by 2.04, 1.81 and 2.47 pp at 10%, 50% and 100% supervision, respectively, leading on 7/10, 8/10 and 10/10 targets. These results suggest that source validation provides an informative basis for selecting ExpARSeg checkpoints whose strengths extend to independent test cases and new domains. This is practically valuable because model selection can rely on a small labeled source subset while yielding broad improvements in target segmentation.

For unexpected results from other methods, a mechanistic interpretation is that these methods learn which cues to trust from the observed data. CausalCLIPSeg's lesion–confounder separation can suppress useful MRI features when anatomy or tissue contrast changes. CAT-Seg's similarity aggregation can reinforce misleading ultrasound matches caused by speckle and weak boundaries. In MedCLIPSeg, validation selection may favor attention variances that assign reliability according to appearances prevalent in a small cohort. Validation DSC therefore rewards the checkpoint whose suppression, aggregation or weighting best fits those cases. On test lesions with different textures or morphology, these operations may suppress lesion evidence or amplify background responses.

Q5. Please assess the statistical reliability via at least three training seeds.

We have also reported results with mean and variance in other responses. The three training seeds are 665,666 and 667, where 666 is also used in the main submission and the other two are supplemented. Here, we also present representative comparisons as follows:

MethodBUSI 10%BUSI 50%
CAT-Seg69.74±0.5160.61±1.47
ExpARSeg (ours)69.76±1.2680.49±0.60
MethodBTMRI 10%BTMRI 50%
CausalCLIPSeg75.94±0.1366.31±2.65
ExpARSeg (ours)82.51±1.9087.90±0.24
MethodMedCLIPSegExpARSeg
BUSI (10%)57.73±2.1569.76±1.26
BUSBRA58.03±2.9358.99±6.14
BUSUC77.68±1.7776.09±3.85
BUID69.98±2.3566.26±3.15
UDIAT67.13±0.7370.19±2.52
MethodMedCLIPSegExpARSeg
BUSI (50%)80.37±0.6680.49±0.60
BUSBRA66.38±6.6474.46±1.79
BUSUC81.54±1.3582.01±2.16
BUID75.74±3.4379.88±1.74
UDIAT72.32±3.8473.57±3.38
MethodMedCLIPSegExpARSeg
BUSI (100%)84.09±0.9285.08±1.10
BUSBRA67.04±3.3174.87±1.18
BUSUC79.71±2.6083.60±1.88
BUID76.80±4.7180.12±1.49
UDIAT72.88±2.6875.29±4.11

In general, ExpARSeg achieves higher mean DSC in most settings, with the larger gains retained in the three-seed averages. This pattern supports its overall performance advantage across annotation budgets and target datasets. A smaller subset of comparisons shows relatively small improvements, including BUSI 10% against CAT-Seg and BUSI 50% against MedCLIPSeg, where the mean differences are modest relative to the observed variation.

Reviewer_3iFqOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • The paper addresses data-efficient medical image segmentation, aiming to improve performance under limited annotations and domain shifts.
  • It proposes ExpARSeg, combining Evidential Pixel-Text Adapters for explicit vision-language adaptation and Selective Spatial Rendering for refining uncertain segmentation regions.
  • Extensive evaluations across four imaging modalities and ten held-out datasets demonstrate improved sample efficiency and domain generalization.
Critical strengths
  • The proposed EPT and SSR modules provide a coherent framework for improving data-efficient medical image segmentation through explicit adaptation regulation.
  • The paper evaluates the method across four imaging modalities and ten held-out datasets, with comprehensive ablation studies demonstrating improved sample efficiency and domain generalization.
Critical weaknesses
  • The proposed evidential states are deterministically derived from pixel-text similarity scores, with foreground probability and uncertainty-related measures expressed as simple functions of these scores. It remains unclear whether the Dirichlet-based formulation provides meaningful advantages over simpler similarity-based gating or uncertainty estimation methods.
  • Appendix A.1 states that the benchmark text prompts were originally constructed using attributes extracted from images, segmentation masks, and labels. Such prompts may contain spatial or semantic information derived from ground-truth annotations, potentially affecting the interpretation of the reported domain generalization results. Although the same prompts are used for all baselines, their availability in practical inference settings remains questionable.
Questions for authors
  • Can the authors compare EPT with direct similarity-based gating, confidence-based attention, or entropy-based uncertainty regulation under identical settings?
  • Can the authors report segmentation performance using generic, class-level text prompts that do not rely on ground-truth mask information, particularly on unseen target datasets?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 3iFq [Q1]: Advantages of Dirichlet-based Formulation

We thank the reviewer for the serious and insightful review, and we hope our responses can deal with the main concerns.

W1. It remains unclear whether the Dirichlet-based formulation provides meaningful advantages over simpler similarity-based gating or uncertainty estimation methods.

We have followed the reviewer's suggestion to compare our EPT with direct similarity-based gating, confidence-based attention, and entropy-based uncertainty regulation under identical settings. We have also included results from two extra random seeds for a more comprehensive evaluation. Results are as follows:

10% Data:

DatasetSimilarity gatingConfidence attentionEntropy uncertaintyExpARSeg
Overall69.32 ± 2.6670.48 ± 0.9270.94 ± 1.1373.07 ± 0.66
BUSI73.84 ± 1.0772.11 ± 1.6971.51 ± 3.4969.76 ± 1.26
BUSBRA55.27 ± 8.9762.04 ± 0.3461.85 ± 4.1058.99 ± 6.14
BUSUC65.16 ± 10.5671.09 ± 1.7171.82 ± 3.6276.09 ± 3.85
BUID43.26 ± 11.6655.16 ± 4.7556.69 ± 8.2166.26 ± 3.15
UDIAT63.30 ± 3.4059.94 ± 6.6163.53 ± 3.1270.19 ± 2.52
BTMRI81.77 ± 0.1482.00 ± 0.4682.18 ± 0.5282.51 ± 1.90
BRISC73.90 ± 0.5973.91 ± 0.3074.54 ± 0.4675.74 ± 1.52
Kvasir-SEG80.44 ± 0.4479.79 ± 2.1380.38 ± 0.3580.67 ± 1.30
BKAI64.20 ± 1.1563.94 ± 3.5164.00 ± 1.4465.49 ± 1.98
CVC-ColonDB57.61 ± 1.1957.03 ± 1.1156.49 ± 0.5758.46 ± 2.61
CVC-ClinicDB67.37 ± 0.6467.09 ± 3.1766.74 ± 0.8368.89 ± 1.55
CVC-30074.82 ± 1.1474.85 ± 1.4274.73 ± 1.9876.61 ± 0.86
ISIC89.65 ± 0.4989.27 ± 0.4489.30 ± 0.2790.56 ± 0.29
UWaterlooSkinCancer79.86 ± 3.9878.52 ± 1.4479.46 ± 1.7682.80 ± 2.11

50% Data:

DatasetSimilarity gatingConfidence attentionEntropy uncertaintyExpARSeg
Overall77.11 ± 1.5876.82 ± 0.4676.58 ± 0.6180.28 ± 0.35
BUSI80.43 ± 2.1780.62 ± 0.4481.32 ± 0.7580.49 ± 0.60
BUSBRA69.36 ± 2.7966.83 ± 1.9467.41 ± 1.6874.46 ± 1.79
BUSUC76.84 ± 5.8576.58 ± 1.1477.66 ± 1.2682.01 ± 2.16
BUID68.36 ± 9.7269.72 ± 5.3970.28 ± 3.6079.88 ± 1.74
UDIAT67.23 ± 4.2163.54 ± 6.3564.74 ± 8.1473.57 ± 3.38
BTMRI86.38 ± 0.2085.99 ± 0.0585.94 ± 0.4687.90 ± 0.24
BRISC78.77 ± 0.7178.10 ± 0.4078.46 ± 0.4080.50 ± 1.52
Kvasir-SEG87.42 ± 0.7487.70 ± 0.1786.90 ± 0.8587.97 ± 0.33
BKAI75.38 ± 1.0774.14 ± 1.4073.24 ± 0.6774.65 ± 2.56
CVC-ColonDB67.26 ± 1.2666.62 ± 2.2565.06 ± 0.9368.10 ± 0.64
CVC-ClinicDB75.14 ± 2.6874.42 ± 2.1173.69 ± 0.9276.53 ± 1.18
CVC-30079.76 ± 1.6581.05 ± 2.0678.92 ± 1.3980.64 ± 1.36
ISIC91.64 ± 0.3491.25 ± 0.4391.60 ± 0.1592.19 ± 0.27
UWaterlooSkinCancer75.61 ± 3.0278.91 ± 0.9776.95 ± 4.9885.09 ± 2.43

100% Data:

DatasetSimilarity gatingConfidence attentionEntropy uncertaintyExpARSeg
Overall80.11 ± 0.4977.78 ± 0.6874.79 ± 4.5681.98 ± 0.42
BUSI84.40 ± 0.1683.58 ± 0.3884.71 ± 0.3085.08 ± 1.10
BUSBRA72.71 ± 1.3866.02 ± 4.5958.32 ± 13.5174.87 ± 1.18
BUSUC81.55 ± 1.2677.67 ± 3.2861.53 ± 17.8583.60 ± 1.88
BUID76.08 ± 0.7671.39 ± 1.5753.29 ± 18.8480.12 ± 1.49
UDIAT69.12 ± 0.9862.70 ± 4.5663.62 ± 7.3475.29 ± 4.11
BTMRI87.94 ± 0.1187.48 ± 0.7487.85 ± 0.2188.98 ± 0.27
BRISC80.50 ± 0.1478.95 ± 1.3480.47 ± 0.4678.48 ± 4.74
Kvasir-SEG89.41 ± 0.2688.83 ± 0.6788.85 ± 0.9689.81 ± 0.15
BKAI78.20 ± 0.5977.65 ± 1.1177.27 ± 1.2679.48 ± 1.20
CVC-ColonDB71.12 ± 1.1670.72 ± 1.1969.85 ± 0.7373.17 ± 0.59
CVC-ClinicDB78.53 ± 0.0279.15 ± 2.5576.58 ± 0.4180.85 ± 0.15
CVC-30082.23 ± 1.2981.74 ± 1.5981.56 ± 1.0783.39 ± 0.67
ISIC92.20 ± 0.1691.97 ± 0.1991.85 ± 0.0792.59 ± 0.04
UWaterlooSkinCancer77.50 ± 5.9871.08 ± 8.7671.37 ± 14.6982.03 ± 2.41

Our method achieves the best mean DSC on 12/14, 11/14, and 13/14 datasets at 10%, 50%, and 100% annotation budgets, respectively. Its overall DSC exceeds the strongest alternative by 2.13, 3.17, and 1.87 pp, with lower standard deviations of overall DSC than those of all three alternatives at every budget. Together, these broad gains and reduced aggregate variability across seeds support the practical advantage of our Dirichlet-based regulation over these simpler alternatives.

AuthorsOpenReview ↗

Response to Reviewer 3iFq [Q2]: Segmentation Performance using Mask-independent Prompts

W2. Although the same Mask-dependent prompts are used for all baselines, their availability in practical inference settings remains questionable.

Our main experiments follow the released MedCLIPSeg benchmark, with all methods receiving the same image-text pairs. This shared setting allows us to compare adaptation methods under a common level of textual assistance. The reported generalization results therefore describe transfer across datasets with benchmark-provided descriptions.

To examine performance with prompts available in practice, we further evaluate performance using instructions specified solely from the segmentation task. For each modality, we write six templates in advance using only the modality and target category, such as “Segment the polyp regions in this endoscopic image.” All methods are retrained with these templates, cycling through them during training and averaging their predictions with equal weights at inference. The templates and weights are fixed across annotation budgets, seeds, and source and target datasets. Their construction and use are independent of patient annotations and validation results. In practice, the requested task and modality determine the prompt pool, allowing the same instructions to be reused for every patient. We compare our method with the strongest baseline MedCLIPSeg as follows:

10% Data:

MethodMedCLIPSegExpARSeg (ours)
Overall68.86±0.9470.84±0.53
BUSI57.25±4.3259.63±3.33
BTMRI78.17±0.8880.35±0.07
Kvasir-SEG77.43±1.1478.62±1.36
ISIC90.18±0.2590.23±0.10
BUSUC73.54±3.7376.28±0.99
BUSBRA52.42±6.5756.36±1.98
BUID64.48±2.8169.32±1.04
UDIAT64.43±3.4667.18±0.55
CVC-ColonDB53.23±1.1855.92±1.09
CVC-ClinicDB66.11±0.6467.45±0.72
CVC-30071.10±3.2274.17±1.63
BKAI62.73±2.0864.83±1.61
BRISC70.82±0.5373.46±0.62
UWaterlooSkinCancer82.17±2.2677.87±2.48

50% Data:

MethodMedCLIPSegExpARSeg (ours)
Overall75.93±0.1677.16±0.31
BUSI67.96±2.9272.61±2.20
BTMRI84.08±0.1985.36±0.26
Kvasir-SEG84.71±0.7586.09±0.36
ISIC91.47±0.1591.76±0.27
BUSUC81.01±1.2682.85±1.17
BUSBRA69.02±1.0469.07±2.29
BUID75.51±0.9476.90±0.75
UDIAT71.38±1.4071.44±3.90
CVC-ColonDB62.12±0.7564.15±0.24
CVC-ClinicDB72.73±1.4773.36±1.22
CVC-30077.90±1.3680.61±0.76
BKAI68.42±0.6869.24±0.43
BRISC72.48±0.2174.45±1.23
UWaterlooSkinCancer84.30±1.3382.36±3.75

100% Data:

MethodMedCLIPSegExpARSeg (ours)
Overall78.53±0.1579.44±0.20
BUSI74.62±2.4078.35±2.00
BTMRI86.06±0.5987.92±0.31
Kvasir-SEG87.44±0.2587.55±0.44
ISIC91.53±0.2291.84±0.06
BUSUC84.71±0.3484.92±0.47
BUSBRA73.15±0.8973.60±1.54
BUID77.31±0.6077.43±1.97
UDIAT74.49±1.6576.07±0.99
CVC-ColonDB66.41±0.3867.24±0.10
CVC-ClinicDB76.59±2.1776.75±1.18
CVC-30080.63±1.5882.03±1.27
BKAI72.24±0.5771.86±0.36
BRISC72.02±2.3175.03±0.55
UWaterlooSkinCancer82.26±0.8281.54±1.08

With task-level prompts, our ExpARSeg outperforms MedCLIPSeg on 13/14, 13/14, and 12/14 datasets at the 10%, 50%, and 100% annotation budgets, respectively. These include all four source datasets at each budget and 9/10, 9/10, and 8/10 target datasets. Overall DSC increases from 68.86 to 70.84, from 75.93 to 77.16, and from 78.53 to 79.44, with the largest improvement at 10% annotations. The gains on most target datasets show that ExpARSeg also transfers well when the text consists of generic task instructions. Its advantage over MedCLIPSeg therefore carries over to a setting where prompts can be supplied directly from the requested segmentation task, without access to patient-specific mask annotations.

AuthorsOpenReview ↗

Follow-up to Reviewer 3iFq

Thank you again for your valuable and constructive review. All your suggestions have been added in the revised paper.

Where we incorporated your suggestions

  • Simpler regulation mechanisms (Q1). ExpARSeg exceeds the strongest of similarity gating, confidence-based attention, and entropy-based regulation by 2.13, 3.17, and 1.87 points in overall DSC at 10%, 50%, and 100% supervision, respectively (App. B.8).
  • Prompts available in practice (Q2). We retrained both methods with six fixed task instructions per modality. ExpARSeg leads MedCLIPSeg on 9/10, 9/10, and 8/10 target datasets at the same budgets (App. B.7).

We appreciate the reviewer for the time and consideration given to our paper.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response: Results for Main Concerns and PDF Revision

General Response to All Reviewers

We thank all three reviewers for their careful and constructive reviews. Here, we outline the concerns shared by most reviewers, along with our general responses. More details can be found in the per-reviewer replies. We have also added the new results into the revised submission.

General Response to Core Concerns

  • Performance with mask-independent prompts (3iFq, RtWf, UYdk). We retrained ExpARSeg and MedCLIPSeg with six mask-independent task templates per modality and equal-weight prediction averaging. Template construction and weighting are independent of patient annotations and validation results. Across three seeds, ExpARSeg leads on 13/14, 13/14, and 12/14 datasets at 10%, 50%, and 100% source annotation budgets, respectively, including 9/10, 9/10, and 8/10 targets. Overall DSC improves by 1.98, 1.23, and 0.91 percentage points (pp), showing that its advantage extends to prompts prepared from the requested segmentation task.
  • Benefits over simpler regulation mechanisms (3iFq, RtWf). We directly compared ExpARSeg with similarity-based gating, confidence-based attention, and entropy-based uncertainty regulation over three training seeds. At 10%, 50%, and 100% source annotations, ExpARSeg exceeds the strongest alternative in overall DSC by 2.13, 3.17, and 1.87 pp and leads on 12/14, 11/14, and 13/14 datasets. Its overall DSC also has a lower seed standard deviation than all three alternatives at each budget. These segmentation results support the practical benefit of the evidential regulation design.
  • Target generalization across annotation budgets and modalities (RtWf, UYdk). With the original benchmark prompts, we report three-seed comparisons with MedCLIPSeg separately for all ten target datasets at 10%, 50%, and 100% source annotations. ExpARSeg improves mean target DSC by 2.04, 1.81, and 2.47 pp and leads on 7/10, 8/10, and 10/10 targets, respectively. MRI and endoscopy show the largest target gains at 10%, while ultrasound and skin show the largest gains at 100%. These results establish broader transfer benefits across annotation budgets and clarify how their magnitude varies by modality.

Changes in Revised PDF

  • App. A.1 (Dataset Settings) clarifies that all methods share the same training, validation, and test manifests at each annotation budget and select checkpoints by the highest source validation DSC.
  • App. B.7 (Detailed Performance with Mask-independent Prompts) adds the protocol using six fixed task templates per modality and a consolidated table of three-seed results at 10%, 50%, and 100% source supervision. The analysis covers all four source and ten target datasets using prompts constructed independently of patient annotations and validation results.
  • App. B.8 (Detailed Benefits over Simpler Regulation Mechanisms) adds three tables comparing similarity-based gating, confidence-based attention, and entropy-based uncertainty regulation with our Dirichlet-based design across annotation budgets. The accompanying analysis reports dataset-wise gains, overall DSC, and variation across training seeds.
  • App. B.9 (Detailed Analysis of Main Comparison) expands the comparison with MedCLIPSeg under the original benchmark prompts to three seeds and all 14 datasets. A consolidated table and accompanying discussion report target performance at each annotation budget and explain how the gains vary across modalities.
  • App. B.10 (Role of TC-KL in Explicit Regulation Mechanism) adds a three-seed comparison of MedCLIPSeg, ExpARSeg without TC-KL, and the full model at 100% source supervision. The discussion distinguishes source and target effects and connects the segmentation results to the existing gradient analysis and qualitative evidence, clarifying the complementary contributions of state regulation and TC-KL.
  • App. B.11 (Effect of SSR on Confidently Negative Regions) adds a three-part table reporting target-dataset recovery rates, selection coverage and recovery among selected pixels by modality, and recovery counts for the existing qualitative examples. The text explains how relative margin ranking admits confident misses for refinement and how visual corrections can reverse their negative predictions.
AuthorsOpenReview ↗

Summary of the Discussion

We thank the Area Chair and all the reviewers for the time and consideration given to our paper. The questions below led to new experiments, extra clarifications, and follow-up analysis, all detailed in our replies and the revised PDF.

Review questions and our responses

  • Performance with Mask-independent Prompts (3iFq Q2, RtWf Q1, UYdk Q1). With both models trained and tested using only such reusable task instructions, our ExpARSeg retains its advantage over MedCLIPSeg, the strongest baseline, on most unseen datasets at every annotation budget (App. B.7).
  • Benefits over Simpler Regulation Mechanisms (3iFq Q1, RtWf Q2). Our evidential regulation gives higher mean segmentation accuracy than direct similarity gating, confidence weighting, or entropy weighting at every annotation budget (App. B.8).
  • Detailed Dataset-wise Analysis (UYdk Q2, RtWf Q6). On the ten unseen targets, ExpARSeg improves average DSC even with limited source labels. The benefit extends beyond ultrasound and skin imaging, with the largest gains at 10% supervision on MRI and endoscopy (App. B.9).
  • Experimental Validity (RtWf Q3,Q4,Q5) . We employ validation-based protocol for checkpoint selection. Three-seed comparisons are used to evaluate the statistical significance of the improvements (App. A.1 and Q3–Q5 reply).
  • Role of TC-KL (RtWf Q7) . ExpARSeg already outperforms MedCLIPSeg overall when trained with segmentation loss alone. TC-KL mainly helps the learned states transfer to unseen datasets (App. B.10).
  • SSR’s Recovery on Confidently Negative Regions (UYdk Q3 and follow-up) . SSR ranks points by absolute margin within each image, so even a confidently missed lesion pixel can be selected. At these locations, fine image details and the text prompt can reverse the coarse prediction. Among lesion pixels initially assigned at least 90% background probability, the recovery rate averaged 8.85% across the ten target datasets (App. B.11).

We thank again for all the constructive and careful reviews, which help make our paper better.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.