Paper Info Reviews Author Feedback Meta-Review Back to top

List of Papers Browse by Subject Areas Author List

Abstract

Airway segmentation from chest CT is essential for safe navigation of the bronchoscopy and the targeting of distal lesions. However, segmentation models frequently exhibit structural failures under domain shift, including disconnected distal branches and boundary leakage. Existing post-hoc refinement methods either apply heuristic topology operations or directly predict corrected masks, implicitly coupling error localization and correction, and limiting controllability. We propose FDRAS (Failure Diagnosis and Repair for Airway Segmentation), a two-stage reliability framework that explicitly decouples failure localization from conservative correction. Given a CT volume and an initial segmentation, FDRAS performs hypothesis verification by predicting two complementary error maps: (1) a Signed Distance Function (SDF)-based geometric error map capturing boundary discrepancies, and (2) a tree-level connectivity error map localizing disconnected branches via skeleton analysis. These error maps condition a mask-only refinement network through cross-attention, enabling targeted updates while preventing topology drift. By externalizing structural inconsistencies and anchoring correction to predicted failure fields, FDRAS enables topology-aware repair rather than direct mask transformation. Under strict zero-shot evaluation across two external datasets and multiple backbones, FDRAS improves branch-level detection and connectivity by up to +4.8 percentage points (pp) compared to the strongest competing refinement, while maintaining competitive voxel-level accuracy, demonstrating its effectiveness as a post-hoc structural reliability layer for deployment.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2460_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/SIRGLab/FDRAS

Link to the Dataset(s)

AirRC dataset: https://figshare.com/articles/dataset/PVAN_data/26878867 Aeropath dataset: https://github.com/raidionics/AeroPath ATM dataset: https://atm22.grand-challenge.org/

BibTex

@InProceedings{ZhaFra_FDRAS_MICCAI2026,
        author = { Zhang, Francis Xiatian AND Dhaliwal, Kevin AND Khadem, Mohsen},
        title = { { FDRAS: Failure Diagnosis and Repair for Airway Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16884},
        month = {September},
        page = {pending}
}

Reviews

Review #1

Please describe the contribution of the paper

FDRAS is a backbone-agnostic post-hoc framework for airway segmentation refinement under domain shift. Given an initial segmentation mask and a CT volume, a diagnosis network predicts two structural error maps: an SDF-based boundary discrepancy field and a skeleton-derived tree connectivity map localizing disconnected branches. A refinement network then applies residual correction to the mask using cross-attention conditioned on these maps. The framework is trained on ATM dataset (240 cases, nnUNet-generated initial segmentations) and evaluated zero-shot on AirRC (254 cases) and AeroPath (27 cases) datasets with two backbone configurations (nnUNet and TfeNet). A single ex-vivo lung specimen provides qualitative validation under an in-vivo-to-ex-vivo distribution shift.
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.

The explicit decoupling of failure localization from correction is the central methodological distinction from prior work. Existing methods either stop at error detection without correction (quality-control methods) or perform correction without prior explicit localization (PTR, UTR, ATRNet). FDRAS places supervised error prediction as a distinct, interpretable stage, giving the refinement network conditioned access to spatially localized failure information rather than operating from global features alone.

The complementary error representation covers the two main failure modes in airway segmentation under domain shift. The SDF error map is dense and continuous, capturing geometric boundary discrepancies throughout the volume. The tree error map is sparse and binary, localizing branch-level connectivity failures. Together they form a structured representation of structural failure that neither metric alone would capture.

The evaluation setup is appropriately strict. Training is on ATM; evaluation is zero-shot on two external datasets with different population characteristics (AirRC: pulmonary nodules; AeroPath: airway pathologies). Testing across two backbones with different architectures and performance levels tests the backbone-agnostic claim under realistic conditions.

The ablation in Table 2 isolates both critical design choices (cross-attention conditioning and error-map conditioning). Removing cross-attention produces the largest structural degradation (TD drop >7 pp on both datasets); removing error conditioning produces a large DSC drop on AirRC (–7.7 pp). Both components are shown to be necessary.

The ex-vivo validation tests a severe distribution shift. The backbone is a 3D U-Net trained on separate private ex-vivo data. The acquisition context shifts from in-vivo CT to ex-vivo CT. FDRAS receives no additional training, yet the predicted SDF error map highlights geometric inconsistencies at distal branches, and the refinement restores branch continuity toward the nodule region. As a single case it is only qualitative evidence, but under conditions where failure would be easy to justify.
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.

With TfeNet on AeroPath, FDRAS reduces DSC by 3.5 pp (87.2 vs 83.7) and precision by 6.6 pp (85.4 vs 78.8). The paper attributes the precision drop to distal branch recovery, not spurious over-segmentation. But a simultaneous DSC drop contradicts this: if the added segments were anatomically correct branches, DSC would be stable or improve. TfeNet already achieves 90.5% TD and 87.4% BD on AeroPath before refinement, suggesting the backbone performs well on this pathological set. FDRAS may be diagnosing structural failures that are not present and introducing incorrect corrections. No per-case analysis or stratification is provided to explain this behavior.

The training configuration for the TfeNet experiments is not stated. The methods section describes FDRAS as trained on nnUNet-generated initial segmentations. If FDRAS was retrained with TfeNet as the source backbone, this must be stated, as it would mean backbone-agnostic operation requires retraining per backbone, a different claim than backbone-agnostic inference.

The diagnosis network is the paper’s most novel component, yet no evaluation of error map prediction quality is provided. The SDF and tree error maps during training are derived from GT. The diagnosis network must learn to predict these from (M, I) at inference without GT. No comparison between predicted and GT-derived error maps is shown. The ablation demonstrates that removing error conditioning hurts performance (Table 2, w/o EM), but this does not confirm the predicted maps accurately localize failures. The corrective benefit may arise from the conditioning mechanism or from generic spatial features of (M, I) rather than from precise error localization.

No confidence intervals or significance tests are reported for either dataset. This could be particularly useful in the case of AeroPath, which contains only 27 cases.

The ablation evaluates error-map conditioning as a whole (w/o EM) but does not separate SDF-only from tree-only configurations. The two maps capture distinct failure modes. Knowing their individual contributions would validate the two-map design and guide future work.
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

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.

(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 decoupled two-stage design is a methodological contribution and the zero-shot evaluation is rigorous. The ex-vivo case is a strong evidence of generalization. The primary uncertainty is the TfeNet-AeroPath regression and the unresolved training configuration for TfeNet experiments, which limit confidence in the backbone-agnostic claim. Rebuttal should address: (a) case-level analysis of the TfeNet-AeroPath regression; (b) an explicit statement of whether FDRAS was retrained for TfeNet. Secondary concerns should be addressed in revision.
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.

My initial review raised five concerns: the TfeNet-AeroPath DSC and precision regression, the unstated TfeNet training configuration, the missing standalone diagnosis-stage evaluation, the absence of confidence intervals on the small AeroPath cohort (n=27), and the missing SDF-only versus tree-only ablation. The rebuttal addresses the main ones. The TfeNet training configuration is now explicit (single FDRAS checkpoint trained on ATM with frozen nnUNet predictions, no TfeNet retraining), the TfeNet-AeroPath regression is given a per-case re-aggregation consistent with distal-branch recovery (TD up in 25/27 cases, BD in 19/27), and FDRAS is honestly reframed as a bounded topology-aware reliability layer rather than a re-segmenter.

The meta-reviewer’s three asks (TfeNet-AeroPath trade-off, diagnosis-stage depth, nnUNet-vs-TfeNet gap) are addressed in the same passes. The standalone diagnosis evaluation and the SDF-only versus tree-only ablation are deferred to the supplement, but the rebuttal’s core answers hold.

I recommend accepting it.

Review #2

Please describe the contribution of the paper

This paper presents FDRAS, a two-stage post-hoc framework for airway segmentation refinement under domain shift. Rather than directly predicting a corrected mask, the method separates failure diagnosis from repair. Given a CT volume and an initial airway segmentation, the diagnosis stage predicts two complementary error maps – an SDF-based geometric discrepancy map and a tree-level connectivity error map derived from skeleton analysis. These predicted failure maps then condition a refinement network through cross-attention, with the aim of making conservative, localised corrections while preserving already-correct regions. The model is trained on ATM and evaluated in a strict zero-shot setting on AirRC and AeroPath, using both nnUNet and TfeNet backbones. The paper reports improvements in structural metrics, including tree detection and branch detection, over prior post-processing baselines, alongside additional qualitative ex-vivo validation.
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.

The paper addresses a clinically relevant and practically important problem. In airway segmentation, structural continuity and distal branch completeness are often more important than voxel-wise overlap alone, particularly for applications such as bronchoscopy navigation and lesion targeting. Framing the problem as post-hoc refinement under domain shift is therefore well motivated.

The central idea of decoupling error localisation from correction is appealing and reasonably original in this setting. The manuscript makes a convincing case that existing approaches tend either to diagnose failures without repairing them, or to repair masks without explicitly modelling the underlying errors. Introducing an intermediate error-representation layer is a sensible and useful contribution.

The structural error representation is also well designed. Combining a dense SDF discrepancy field with a sparse tree-connectivity map aligns well with the failure modes described in the introduction, namely both boundary errors and disconnected distal branches. This is a stronger formulation than relying on connectivity cues alone.

Training on ATM and testing zero-shot on two external datasets is appropriate for the paper’s domain-shift claims, and the use of two different backbones supports the claim that the framework is not tied to a single upstream segmenter. The reported gains on structural metrics appear meaningful, and the ablation results support the contribution of both explicit error conditioning and cross-attention. Overall, the paper is clearly written and easy to follow.
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 main weakness is that the empirical validation remains somewhat narrow relative to the broader framing around deployment reliability. Although the zero-shot evaluation is valuable, the experiments are confined to airway segmentation alone. As a result, the paper is convincing as a task-specific airway refinement study, but somewhat less so as a more general reliability-layer contribution.

A further limitation is that the diagnosis stage, which is conceptually central to the method, is not evaluated in sufficient depth on its own. The paper supervises training using ground-truth discrepancy maps, including the skeleton-derived tree error map, but provides limited analysis of how accurately these maps are predicted at inference or how diagnosis quality relates to final repair quality. Stronger standalone evaluation of this stage would materially strengthen the paper.

Some of the performance trade-offs also warrant more cautious discussion. In Table 1, precision drops noticeably in some settings, particularly with TfeNet on AeroPath. The authors suggest this reflects recovery of distal branches rather than implausible over-segmentation, which is plausible, but the support provided is largely qualitative. A clearer discussion of when such added branches are beneficial versus potentially false positive would improve the paper.

Finally, the ex-vivo validation is interesting but remains qualitative, which limits the strength of any conclusions drawn from it. The paper would also benefit from more explicit analysis of failure cases and limits, particularly since the conclusion notes that severely collapsed segmentations cannot be fully recovered. For a post-hoc repair framework, understanding where the method stops helping is central to its practical value.
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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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

The paper would be strengthened by explicit evaluation of the diagnosis stage itself, for example by quantifying how well the predicted SDF and tree-error maps match their supervised targets, and whether improved diagnosis translates into improved repair.

More detail on the retraining protocol for PTR, UTR, and ATRNet would also be helpful, given the importance of these comparisons. The precision drop observed in some settings deserves fuller discussion, ideally with examples distinguishing acceptable branch recovery from clear false positives.

The ex-vivo section is promising, but should be framed more explicitly as qualitative external validation to avoid overstatement. Since the paper emphasises conservative correction, it would also be valuable to quantify how much of the mask is altered during refinement, or how consistently the method preserves already-correct regions.
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?

This is a solid and clinically relevant paper with a clear motivation, a sensible and reasonably novel method, and convincing zero-shot improvements on structural airway metrics across two external datasets and two backbones. The main contribution is not merely incremental performance improvement, but a useful reformulation of segmentation repair as diagnosis-conditioned correction. The weaknesses relate primarily to depth of analysis rather than to a fundamental flaw: the diagnosis stage is not evaluated as thoroughly as it could be, the ex-vivo evidence is qualitative, and some precision trade-offs deserve more careful discussion. Overall, however, the paper makes a meaningful contribution to airway segmentation reliability and is likely to be of interest 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.

After reading the authors’ rebuttal, I remain supportive of acceptance. The response does not fully resolve every concern raised in my original review, but it addresses the main points in a thoughtful and credible manner and improves confidence in both the framing and the empirical interpretation of the work. In particular, the additional clarification regarding the TfeNet-AeroPath setting is helpful, as the reported per-case branch-level improvements suggest that the method is consistently recovering structural completeness rather than producing only unstable average gains. Overall, the rebuttal strengthens the paper sufficiently for me to retain an accept recommendation, while still recognising that some of the claims should be framed carefully in the final version.

Review #3

Please describe the contribution of the paper

Incomplete airway segmentations render them not very useful in clinical applications. This work contributes an airway segmentation refinement method that uses two steps to identify and fix these problematic segmentations. The work presents a failure diagnosis and repair (FDRAS) framework that shows competitive performance on zero-shot applications, and outperforms other refinement methods on two datasets when starting with initial segmentations from two backbone networks.
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.
- Improving reliability of airway segmentation using a diagnostic method that also repairs errors in distal branches/discontinuities, instead of another segmentation framework is a useful contribution in this work. Many airway segmentation methods need refinement steps to obtain fully useful airway networks.
- Decoupling the errors into boundary discrepancies (measured using signed distance function (SDF)) and branch connectivity using skeleton analysis, is an interesting way to improve the reliability of airway segmentation quality.
- Strong zero shot performance is reported in Table 1, and thorough evaluations (including ablations) are reported.
- Additional validation on ex-vivo human lung specimen is a valuable contribution that is seldom seen in airway segmentation works. I hope the authors will make this specimen data available along with the code.
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.
- Method is not described in detail. For example, the authors say “feature embeddings extracted by lightweight 3D convolutional blocks” or “employ a spatially reduced cross-attention operator”. The specific choice for these modules is not defined anywhere. So, would any “lightweight 3D convolution” block work? And the specific value for the reduction factor s is not mentioned.
- The initial segmentations analysed are shown using two backbone networks (nnUnet and TFeNet). Why are the refinement approaches, including FDRAS unable to improve nnUnet to the level of TFeNet? Is this because there are large gaps in performance already with the backbones (about 12 pp DSC)? This points to the limitations of refinement approaches as they are heavily dependent on the initial segmentation. What is the prescribed approach then to obtain useful segmentations?
- The qualitative comparison shown in Fig 2 is not actually showing a problem with distal airways. The zoom inset is for the main bronchi, and I am quite surprised that already the input segmentation has the “hole” which is not rectified by other methods. Why is this so challenging? Perhaps focusing on loss of connectivity in smaller airways might make this qualitative example more compelling.
- Justification for chosen hyperparameters (five lambda terms). How were these parameters chosen? And how sensitive is the performance of the method to these choices? Also, curious to see many of these parameters are simply set to 1? Is it fixed across different datasets?
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

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.

(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?

I have a few concerns about the choice of hyperparameters, model design, and usefulness of refinement approaches. These can be clarified in the rebuttal.
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.

My concerns raised about the improvement on initial segmentation were partially addressed; the authors also commit to including more method details in the revised version.

Author Feedback

We thank the AC and reviewers for their feedback and address the main concerns: TfeNet-AeroPath trade-off, diagnosis, backbone, and protocol clarity.

1.TfeNet-AeroPath Regression and Precision Drop (AC, R1 & R2) We agree that the TfeNet-AeroPath results require careful interpretation. Reorganizing the existing case-level outputs (without new experiments, per MICCAI rules) shows that FDRAS consistently improves branch-level metrics: TD rises in 25/27 cases and BD in 19/27, while Dice improves in roughly half of the cases. Voxel-wise DSC and precision may decrease slightly relative to the ground truth, as FDRAS occasionally alters voxels outside annotated branches. While this enhances distal connectivity, these modifications are not explicitly validated and are noted as a limitation and a potential direction for future research. Branch-level gains indicate that FDRAS effectively restores distal structures without introducing grossly implausible anatomy. If permitted, we will include this per-case summary with relevant statistical reporting in the supplement.

2.Diagnosis-Stage Evaluation (AC, R1 & R2) In the submitted manuscript, diagnosis is evaluated through its functional role in repair: it is supervised with GT-derived SDF and skeleton/tree error targets, and removing error-map conditioning reduces DSC by 7.7 pp on AirRC and 6.5 pp on AeroPath. This supports the usefulness of predicted error maps for repair, but we agree it is not a complete standalone diagnosis benchmark. If permitted, we will add map-level diagnosis assessment from the existing predictions and GT-derived targets in the supplement, along with SDF-only/tree-only ablations.

3.Dependence on Initial Segmentation Quality (AC & R3) The gap between nnUNet and TFeNet arises from initial segmentation quality: FDRAS improves structural completeness but cannot create branches entirely missing in the initial mask. Nevertheless, applying FDRAS to nnUNet narrows the gap with TFeNet: DSC on AirRC differences decreases from 12.6 pp to 4.1 pp, while TD/BD differences also decrease, indicating meaningful structural recovery. On AeroPath, similar trends are observed. By targeting errors without full re-segmentation, FDRAS can improve the structural completeness of the backbones evaluated. We will clarify this in the Conclusion and present FDRAS as a bounded, topology-aware reliability layer for targeted correction, designed to enhance existing segmentations rather than replace them.

4.Backbone Protocol and Implementation Clarity (R1 & R3) R1 asked about backbone handling: the same FDRAS checkpoint trained on ATM with frozen nnUNet predictions was used for all nnUNet and TFeNet evaluations; no TFeNet-specific retraining was performed.

R3 asked for architectural details: lightweight 3D blocks use two 3×3×3 convolutions with instance normalization and ReLU, with up-sampling via ConvTranspose3d and skip connections. Cross-attention uses spatial reduction s=4 with residual addition. Loss weights were set to default values to avoid scenario-specific tuning. These details clarify model design and hyperparameters. Code and weights will be released upon acceptance; architectural details will be added to Methods, and sensitivity analysis noted as future work.

5.Ex-Vivo Validation and Qualitative Claims (R1, R2 & R3) R3 asked why standard methods fail and whether Fig. 2 shows distal airway failure. We will revise the figure/caption to highlight a clearer distal case. The zoom inset shows a “hole” in the main bronchi not corrected by other methods, which assume common in-vivo shapes and do not model structural errors, limiting recovery of ex-vivo cases. We agree the validation is qualitative: FDRAS can operate on extreme cases, while main evidence remains zero-shot evaluation on AirRC and AeroPath.

We thank the AC and reviewers again and will revise the manuscript accordingly.

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.

Although three reviewers give weak accept score, they raise several major concerns, which need to be further evaluated in the rebuttal. These major concerns include:

1) Potential ineffective of FDRAS module: “With TfeNet on AeroPath, FDRAS reduces DSC by 3.5 pp (87.2 vs 83.7) and precision by 6.6 pp (85.4 vs 78.8). The paper attributes the precision drop to distal branch recovery, not spurious over-segmentation. But a simultaneous DSC drop contradicts this: if the added segments were anatomically correct branches, DSC would be stable or improve. TfeNet already achieves 90.5% TD and 87.4% BD on AeroPath before refinement, suggesting the backbone performs well on this pathological set. FDRAS may be diagnosing structural failures that are not present and introducing incorrect corrections.”

2) The diagnosis stage, which is central to the method, is not evaluated in sufficient depth. “The paper provides limited analysis of how accurately these (discrepancy/error) maps are predicted at inference or how diagnosis quality relates to final repair quality.”

3) Limited performance improvement of the refinement method: “The initial segmentations are shown using two backbone networks (nnUnet and TFeNet). Why are the refinement approaches, including FDRAS unable to improve nnUnet to the level of TFeNet? Is this because there are large gaps in performance already with the backbones (about 12 pp DSC)? This points to the limitations of refinement approaches as they are heavily dependent on the initial segmentation.”

Hence, rebuttal is required for this paper.
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.

After the rebuttal, all reviewers keep their initial positive score, although one reviewer’s concern is partially addressed (regarding the improvement on initial segmentation). Recommend acceptance of this work.

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.

All three reviewers recommend acceptance after rebuttal. Reviewers generally agreed that the diagnosis-conditioned repair framework is well motivated, clinically relevant, and demonstrates consistent improvements in structural airway metrics under domain shift. The rebuttal successfully clarified the TfeNet-AeroPath trade-off, the backbone-agnostic evaluation protocol, and the intended role of FDRAS as a topology-aware reliability layer rather than a full re-segmentation method. While a more detailed standalone evaluation of the diagnosis stage would further strengthen the work, the main concerns have been adequately addressed. The presentation could also be improved, including figures, tables, and some implementation details. I recommend acceptance.

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.

The rebuttal addressed the most decision-critical issues. It clarified that a single FDRAS checkpoint trained on ATM with frozen nnUNet predictions was used for both nnUNet and TFeNet evaluations, with no TFeNet-specific retraining. Overall, the reviews and rebuttal support acceptance. The paper makes a meaningful contribution to airway segmentation reliability and is likely to be of interest to the MICCAI community.

back to top

FDRAS: Failure Diagnosis and Repair for Airway Segmentation

Author(s):