List of Papers Browse by Subject Areas Author List
Abstract
Standard Test-Time Adaptation (TTA) methods typically treat inference as a blind optimization task, applying generic objectives to all or filtered test samples. In safety-critical medical segmentation, this lack of selectivity often causes the tumor mask to spill into healthy brain tissue or degrades already-accurate predictions. We propose Hypothesis-Driven TTA, a novel framework that reformulates adaptation as a dynamic decision process. Rather than forcing a single optimization trajectory, our method generates intuitive competing geometric hypotheses: compaction (is the prediction noisy? trim volumetric artifacts) versus inflation (is the valid tumor under-segmented? safely inflate to recover). An unsupervised representation-guided selector then identifies the safest outcome via intrinsic texture consistency. Additionally, a pre-screening Gatekeeper prevents negative transfer by skipping adaptation on confident cases. We validate this proof-of-concept on a cross-domain binary brain tumor segmentation task, applying a model trained on adult BraTS 2023 gliomas to unseen pediatric and more challenging meningioma target domains. HD-TTA improves safety-oriented outcomes (95th percentile Hausdorff Distance (HD95) and Precision) over several state-of-the-art representative baselines. Under severe domain shift (BraTS-MEN), HD-TTA demonstrates its robustness by reducing HD95 by approximately 6.4 mm and improving Precision by over 4%, while maintaining comparable Dice scores. These results demonstrate that resolving the safety-adaptation trade-off via explicit hypothesis selection is a viable, robust path for safe clinical model deployment. The code is available at https://github.com/kartikvi01091896/HD-TTA.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3549_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/kartikvi01091896/HD-TTA
Link to the Dataset(s)
BraTS 2023 GLI dataset: https://www.synapse.org/Synapse:syn51514105
BraTS 2023 PED dataset: https://www.synapse.org/Synapse:syn51514108
BraTS 2023 MEN dataset: https://www.synapse.org/Synapse:syn51514106
BibTex
@InProceedings{JhaKar_HDTTA_MICCAI2026,
author = { Jhawar, Kartik AND Wang, Lipo},
title = { { HD-TTA: Hypothesis-Driven Test-Time Adaptation for Safer Brain Tumor Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose Hypothesis-Driven Test-Time Adaptation (HD-TTA), a novel unsupervised inference framework designed to improve the safety of brain tumor segmentation under severe domain shifts. Unlike standard TTA methods that blindly apply generic optimization objectives to all test samples, HD-TTA introduces a “Gatekeeper” to bypass confident predictions. For uncertain cases, it dynamically generates competing geometric hypotheses (compact denoising vs. diffuse recovery) and selects the safest refinement using an intrinsic representation-guided selector.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Strong Clinical Motivation: Reformulating test-time adaptation from a blind optimization task to a dynamic decision process directly addresses the critical issue of “negative transfer” in medical imaging, where unnecessary updates degrade already-accurate predictions.
2.Appropriate Evaluation Metrics: The emphasis on safety-critical metrics like Hausdorff Distance (HD95) and Precision, rather than relying solely on Dice scores (which often mask boundary failures and false positives), aligns perfectly with clinical deployment needs.
3.Robust Out-of-Distribution Evaluation: The experimental design impressively validates the framework under severe domain shifts—training on adult gliomas (BraTS-GLI) and testing on pediatric cases (BraTS-PED) and meningiomas (BraTS-MEN)—without target-specific hyperparameter tuning.
4.Thorough Ablation Study: The ablation experiments clearly validate the necessity of each proposed component (the Gatekeeper, the specific hypotheses, and the edge map constraint), proving that the selector effectively identifies and rejects unsafe parameter updates.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Simplification to Binary Segmentation: The framework is evaluated strictly as a binary tumor segmentation proof-of-concept. Given that the BraTS datasets provide rich, multi-class labels (edema, enhancing core, necrosis), limiting the task to whole-tumor segmentation bypasses the complex inter-class boundary dynamics that a clinical model must navigate.
2.Heuristic Threshold Dependency: The method relies on several hard-coded, heuristically derived thresholds (e. g. , Gatekeeper flagging at < 300 voxels or uncertainty > 5%, and the selection threshold $S_{rep} > 0.95$). While fixed across this study, it is unclear how sensitive the framework is to these specific values across different modalities or scanner protocols.
3.Computational Overhead: The 1000 logit optimization steps for flagged cases introduce a notable latency (~21.9 seconds per case). While acceptable for offline analysis, this computational overhead might hinder integration into real-time clinical workflows compared to lighter TTA methods.
4.Baseline Implementation Anomaly: The authors note that the TEGDA baseline produced “unchanged output masks” in their nnU-Net v2 configuration. This raises concerns about whether this specific baseline was optimally configured or fairly represented, as it effectively defaulted to a non-adapted state.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
As a reviewer, my goal is to help strengthen your manuscript. This is a very well-written paper that addresses a highly relevant issue in TTA. To strengthen the methodology, I recommend including a brief sensitivity analysis of your core hyperparameters (e. g. , the $\gamma$ tolerance scaling factor and the $S_{rep}$ threshold) to demonstrate that your “safe operational boundaries” are robust and not overly tuned to the BraTS feature space. Please clarify the implementation details regarding the TEGDA baseline failure. If a state-of-the-art baseline fails to update in your pipeline, providing a deeper technical explanation as to why the architectural incompatibility occurs would be beneficial for transparency. While the binary proof-of-concept is acceptable, a brief discussion on how the hypotheses generation might mathematically scale to multi-class scenarios (e. g. , handling competing growth between adjacent edema and necrosis boundaries) would greatly enhance the paper’s theoretical depth.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The manuscript presents a highly original, clinically motivated approach to test-time adaptation. By prioritizing boundary safety and false-positive suppression over marginal overlap gains, the authors tackle a genuine barrier to the clinical deployment of AI. The methodology is sound, the validation across severe domain shifts is compelling, and the ablation study effectively isolates the contribution of each module. Despite the limitation of being a binary proof-of-concept with some computational overhead, the core conceptual contribution is strong and highly relevant to the MICCAI community.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors provided a comprehensive and convincing rebuttal that thoroughly addressed my initial concerns.
First, the clarification regarding the TEGDA baseline anomaly—specifically the architectural incompatibility between TEGDA’s batch normalization updates and the nnU-Net v2’s Instance Normalization—completely resolves my reservations about the baseline comparisons. This technical transparency is appreciated and should be explicitly noted in the final manuscript or supplementary material.
Second, the defense of the fixed thresholds (e.g., the Gatekeeper parameters and S_rep > 0.95) is well-argued. Establishing these boundaries geometrically a priori, and proving their robustness across severe domain shifts without target-specific tuning, effectively counters the critique of heuristic dependency.
Finally, the provided roadmap for multi-class adaptation (using class-conditional S_rep and modality-specific geodesic barriers) demonstrates the theoretical scalability of the framework. The computational overhead is also adequately justified given the targeted routing of the Gatekeeper (flagging only 23.6% of cases in the more stable PED domain).
The core contribution remains highly relevant to the clinical translation of AI in medical imaging, and the authors have robustly defended their methodological choices. I maintain my positive assessment and recommend acceptance.
Review #2
- Please describe the contribution of the paper
This paper proposes Hypothesis-Driven TTA (HD-TTA). The main contributions of the paper include: (1) A Gatekeeper to select the confident samples and prevent unnecessary updates; (2) A Hypothesis-Conditioned Refinement module for adaptation and (3) A Representation-Guided Selection module to output the safest correction.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.As the authors mentioned in the manuscript, the proposed hypothesis-driven refinement framework is general and can be extended to multi-class or other structured prediction tasks, which is feasible in different applications. 2.This paper mentioned geometric hypotheses (compaction versus inflation) to guide the adaptation, which can inspire the works focused on brain tumor segmentation.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Writing and Formatting Issues: (1) Inaccurate abbreviations: In the abstract, the full metric name for HD95 should be explicitly stated as the “95th percentile of Hausdorff Distance,” rather than simply “Hausdorff Distance.” (2) Poor structural organization: The paper redundantly mentions that the backbone is an nnU-Net v2 pre-trained on BraTS-GLI across the Introduction, Method, and Experiment sections. It is recommended to consolidate all experimental setup details exclusively into the Experiment section to reduce clutter; this would also better highlight the method’s robustness. Similarly, the specific numerical values for various hyperparameters should be detailed in the experimental section rather than being mixed into the method description. (3) Vague method descriptions: The method introduces a “gravity loss V(P)” but fails to provide an explicit formula for its calculation. Additionally, terms such as the “inflation term” and “g” are poorly defined.
2.Lack of Theoretical Justification: The “Gatekeeper” in the proposed method relies on a series of criteria to filter confident samples (e.g., voxel numbers < 300, 0.3 < uncertainty ratio < 0.7). However, the paper provides no theoretical support, relevant references, or hyperparameter-tuning experiments to justify these specific thresholds. Similarly, the threshold setting for the tumor core ($P_0 > 0.8$) in the Representation-Guided Selection lacks adequate justification. Thus, a more clear theoretical justification is needed.
3.Unreasonable Experimental Settings: The BraTS 2023 MEN dataset provides 1,000 cases in its training split, yet the authors randomly sampled only 144 cases as the target domain without clarifying their motivation. It’s recommended to explain why the entire dataset was not utilized as the target domain.
4.Missing Analysis of Weighting Coefficients: In equations (1) and (2) of the Method section, several weighting coefficients are introduced (e.g., $\lambda_{ent}$, $\lambda_{TV}$). However, the authors fail to conduct experimental analyses or ablation studies to evaluate the individual impact of each coefficient, which weakens the persuasiveness of the proposed method.
5.Results Do Not Support Practical Application: Although HD-TTA demonstrates improvements in HD95 and Precision on the BraTS-MEN dataset, the average Dice score across all methods on this dataset is only between 2% and 14%. Such exceedingly poor absolute performance provides minor meaningful guidance for real-world deployment. And in the context of these exceptionally low baseline results, the performance gains introduced by HD-TTA appear highly marginal. So the authors might try to use another dataset for more insightful experimental results.
- Please rate the clarity and organization of this paper
Poor
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper lacks sufficient theoretical justification in both its methodological descriptions and experimental settings. This significantly reduces the overall credibility of the work and increases the difficulty of reproducing the results. Besides, the experimental results fail to effectively demonstrate the superiority of the proposed method.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
Although the authors stated in their response that “All thresholds were geometrically set before evaluation,” they still failed to explain the specific rationale behind this configuration or provide supporting literature. Furthermore, their response regarding TEGDA is inaccurate; both the original TEGDA paper and its accompanying code confirm that TEGDA updates all model parameters—not merely the Batch Norm statistics. Consequently, the authors remain unable to demonstrate that the TEGDA baseline was optimally configured or fairly represented.
Review #3
- Please describe the contribution of the paper
This work presents a novel Hypothesis-Driven Test-Time Adaptation (HD-TTA) framework for the purposes of volumetric medical image segmentation. The framework selectively refines test segmentation outputs, first using a gatekeeper for selective adaptation to mitigate negative transfers on accurate predictions based on the size of a lesion, and or uncertainty ratio. This is followed by generating competing geometric hypothesis to address over-, under-segmentation and noise islands. Finally, an unsupervised selector is used to choose the most probable refinement, based on some intrinsic consistency signal.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
-
HD-TTA carefully attempts to address the real issue of negative transfers in TTA by introducing the gatekeeper mechanism within the framework.
-
This work proposes a noteworthy optimization framework in TTA based on hypothesis and selection.
-
The method is well evaluated under a realistic cross-domain scenario, using various models trained and or tested on the BraTS datasets for pediatrics and adult gliomas, and meningiomas.
-
The use of a HD95 and precision as evaluation alongside dice is well-motivated.
-
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
-
The formulation of the hypothesis can be said to be lacking, as it only addresses two fixed handcrafted features in compaction and diffusion, and no way of identifying diagnostic specific failure. The formulation is constrained to post-hoc selection of these options rather than actually being an adaptive hypothesis framework.
-
The unsupervised selector being based on a simple intensity-consistency assumption seems to be more of a heuristic, depending on manually selected thresholds. This may not hold in heterogeneous setting where tumor labels are not merged into the whole tumor.
-
The gatekeeper mechanism relies on confidence, which may be unreliable in modern deep segmentation architectures (like nnU-Net) that tend to ber overly confident in their predictions. Confidently wrong segmentations may bypass the gatekeeper entirely, where as correct segmentations under ambiguity will undergo unnecessary further TTA.
-
There is a lack of detail in the implementation of the unsupervised selector especially in 3.1 (implementation details).
-
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The idea of hypothesis driven TTA is interesting and potentially impactful. Extending beyond two fixed hypothesis, and making it truly adaptive is sure to strengthen the contribution of this work. The authors should further consider how the gatekeeper mechanism operates, as confidence (uncertainty ratio) may not be the best way forward under uncalibrated models for a ‘plug-in’ method.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The authors present a good case for why hypothesis-driven TTA is necessary. The proposed framework has various components with stand-alone shortcoming worth addressing in future works (or here, if possible).
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors diligently address the stated concerns in a satisfactory manner.
Author Feedback
Thanks to all reviewers & AC. R1(accept) & R3(weak accept) note novelty/relevance. Responses below: 1) Heuristic thresholds & coefficient ablations [R1, R2, R3]: All thresholds (Gatekeeper: <300 vx, >5% unc.; Selector: S_rep>0.95; core P0>0.8) & loss weights were geometrically set before evaluation (not heuristically via per-domain tuning) & fixed across PED/MEN (Sec. 3.1, fn. 1). Losses are proxy instantiations described as replaceable (Sec. 1). Table 2 tests key decisions: forcing only H_diffuse increases HD95 to 9.31mm; removing Gatekeeper or edge map worsens Dice/HD95.Thus, we do not claim universal optimality, but that a fixed, untuned decision policy improves safety under cross-domain shift. Happy to provide sensitivity & coefficient-level ablations, but Table 2 already proves this safety gain.
2) TEGDA baseline [R1]: TEGDA updates batch norm statistics. nnU-Net v2 uses Instance Norm; per-sample stats recompute on-the-fly, halting persistent update propagation through inference graph (Sec. 3.2). Also, significance is computed against TCA (the strongest baseline), not TEGDA.
3) Post-hoc selection vs. genuine adaptivity [R2, R3]: All TTAs use assumptions (self-training: valid pseudo-labels; alignment: distribution match). HD-TTA’s decisions are instance-specific: the Gatekeeper makes a per-sample go/no-go (Sec. 2.1); the selector chooses hypotheses label-free (Sec. 2.3). Restricting hypotheses to clinical failure modes is deliberate: unconstrained label-free generation would be intractable & unsafe. Our contribution is the label-free per-instance selection principle (Sec. 4). For multi-class extension: K pairs {H_k^compact, H_k^diffuse} with class-conditional S_rep^k, inter-class exclusivity constraints, & modality-specific geodesic barriers (T1-Gd for enhancing core, FLAIR for edema).
4) Selector details [R3]: Full specification is in Sec. 2.3: core P0>0.8, expansion region Δ=H_diffuse\Core, S_rep by Eq. 3, γ=1.5, threshold S_rep>0.95 (fn. 1). In multi-class settings, class-conditional S_rep^k operates on per-class cores, avoiding the merged-label assumption. The intensity-consistency assumption is deliberately conservative; design defaults to H_compact when expansion pixels deviate from the high-confidence core, failing safe. Table 2 validates this: Full HD-TTA matches only-H_compact (HD95 5.35mm). The selector is modular & replaceable (Sec. 2.3).
5) Gatekeeper reliability [R3]: The Gatekeeper uses the intuitive uncertainty ratio (foreground voxels where 0.3<P<0.7), not max-logit confidence. A confidently-wrong network with crisp probabilities yields low uncertainty, skipping adaptation and preserving the baseline. The w/o Gatekeeper ablation confirms: adapting every case worsens Dice & HD95.Alternative non-confidence based (shape plausibility/outlier detection) signals are viable; the Gatekeeper is modular (Sec. 1).
6) BraTS-MEN dice [R2]: Universally low Dice (2-14%) reflects catastrophic cross-domain failure inherited by all baselines, not just HD-TTA (glioma-to-meningioma gap). The MEN experiment is an explicit stress test: asking whether TTA can make a severely shifted model safer. HD95 & Precision govern clinical harm in this regime (Sec. 3.2). HD-TTA’s significant HD95 reduction (6.4mm) & Precision gain (>4%) over TCA prove meaningful harm reduction despite partial recovery.
7) 144 MEN cases [R2]: The 144-case random sample matches BraTS-PED’s test scale for directly comparable metrics. The fixed zero-tuning procedure applies at any scale. This omitted motivation does not affect the reported comparisons.
8) Compute [R1]: Adaptation applies to flagged cases only - 23.6% of PED (5.17s avg.). The 21.74s figure applies under catastrophic MEN shift (99.3% flagged), where safety gains are most critical & offline reporting applies.
9) Writing [R2, R3]: Formulas omitted for space: V(P)=Σ_i P_i‖x_i−μ‖²; g=exp(−α‖∇I‖); inflation=(−1/N)Σ_i P_i. Other presentation edits acknowledged; they do not affect claims.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper presents HD-TTA an unsupervised framework that reformulates adaptation as a dynamic decision process. In particulare HD-TTA is composed by a representation-guided selector then identifies the safest outcome via intrinsic texture consistency. Additionally, a pre-screening Gatekeeper prevents negative transfer by skipping adaptation on confident cases. Reviewers appreciated the strong clinical motivation around negative transfer, the safety-oriented metrics, the robust cross-domain evaluation without target-specific tuning and the ablations validating each component. However, the reviewers identified weaknesses including the simplification to binary tumor segmentation that bypasses multi-class BraTS dynamics [R1], the dependency on hard-coded heuristic thresholds without sensitivity analysis [R1, R2, R3], computational overhead from 1000 logit-optimization steps [R1], an anomalous TEGDA baseline that produced unchanged outputs and weakens the comparison [R1], lack of theoretical justification and explicit formulas for several method components [R2], unjustified experimental setup using only 144 of 1,000 BraTS-MEN cases [R2], missing analysis of weighting coefficients [R2], very low absolute Dice on BraTS-MEN that limits practical relevance [R2], a constrained two-hypothesis design that is not truly adaptive [R3] and unreliable confidence-based gating with modern over-confident networks such as nnU-Net [R3]. The AC read the paper and the reviews and notes some mixed reviews; the safety-oriented framing is interesting, but theoretical justification, threshold sensitivity, and baseline fairness need to be addressed in a rebuttal.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper received mixed reviews and it was invited for rebuttal. The authors submitted a rebuttal to address the reviewer’s comments. After the rebuttal two reviewers gave a positive and one a negative score. The reviewer with the negative scoring highlighted that different components need more intuition and that the discussion about the TEGDA is inaccurate. The AC read the paper, reviews, rebuttal, and final justification, and agrees that the paper has merit; it will benefit from more explanations and discussions. The AC thinks that the paper could make a nice contribution to the conference and encourages the authors to integrate all the answers and the extra questions of R2 in their camera ready version.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Most reviewers find this paper novel and worthy of discussion, with its connection to clinical use being a particular highlight—both of which make it relevant to MICCAI. I partially agree with the dissenting reviewer, however, that the paper remains more of a proof-of-concept and requires further clinical validation for broader impact. Additionally, the authors should clarify the fairness of their baseline implementations.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Though there is still some disagreement between reviewers about the justification for the set thresholds and the proper tuning of TEGDA, the paper falls above the acceptance thresholds for its theoretical and practical metrics, and as most concerns were addressed int he rebuttal.