acceptodds
Under review as a conference paper at ICLR 2027

EndoTME: Physics-Inspired Degradation-Aware Trajectory Memory for Endoscopic Video Enhancement

Abstract

Endoscopic video enhancement is crucial for improving surgical visibility and supporting reliable intraoperative visual perception. However, existing video restoration methods commonly rely on generic temporal feature propagation, which suffers from two key limitations. First, they tend to indiscriminately aggregate historical frame information that may already be distorted or degraded. Second, they lack explicit awareness of endoscopy-specific degradations, making them less effective in challenging surgical scenes involving blood haze, specular reflection, motion blur, low contrast, and temporal instability. To address these issues, we propose EndoTME, an endoscopic video enhancement framework that improves the reliability of temporal information and the adaptivity to complex degradations. Specifically, we develop a reliable trajectory memory retrieval mechanism to selectively exploit trustworthy historical features while suppressing unreliable temporal cues caused by severe image degradation or unstable motion. Such reliability-aware memory selection reduces the cross-frame propagation of corrupted temporal information and improves the temporal coherence of enhanced videos. Meanwhile, we introduce a degradation conditioning mechanism based on physics-inspired visual cues, which explicitly characterizes typical endoscopic degradations and facilitates adaptation of the restoration process to different local degradation patterns. Experiments on a paired synthetic dataset and two real-world endoscopic video datasets demonstrate that EndoTME outperforms representative image and video restoration baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Endoscopic Video Enhancement, Video Restoration, Degradation-Aware Restoration, Temporal Feature Propagation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_CasNOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents EndoTME, a video enhancement framework designed for endoscopic scenes with spatially and temporally varying degradations. The method contains two main components. Reliable Trajectory Memory Retrieval (RTMR) traces historical features along motion trajectories but, rather than aggregating all aligned observations, ranks them using appearance similarity, local quality, flow reliability, and degradation risk before retaining the top memories. Physics-Inspired Degradation Conditioning (PIDC) constructs analytical cue maps for red veiling, specular reflection, blur, weak contrast, structural response, and motion, and uses them at both global and spatial levels to condition restoration and route features through three restoration experts. The method is evaluated on the paired STSVD dataset and two real-world datasets, LSVD and the internally collected EndoReal dataset. EndoTME reports the best results among the compared methods on the paired full-reference metrics and on the reported no-reference and temporal metrics for the real datasets. The paper also includes module-level and internal ablations for RTMR and PIDC, a Top-K sensitivity analysis, qualitative visualizations, and a runtime/parameter comparison.

Critical strengths

A main strength of the paper is the distinction between temporal correspondence and temporal reliability. In endoscopic videos, an aligned historical observation can still be severely corrupted by blood, reflection, blur, or motion-estimation error. RTMR explicitly evaluates historical candidates before aggregation rather than assuming that successful motion alignment makes them useful. The internal ablation supports this motivation: progressively adding feature similarity, local quality, flow reliability, and degradation risk improves the overall restoration and temporal metrics. The experimental study is also reasonably broad. The authors evaluate on a paired benchmark as well as two real-world datasets, including a relatively large internally collected dataset with patient-disjoint splitting. The factorial ablation of RTMR and PIDC is useful, and the reported performance is not obtained at the cost of a particularly large or slow model: EndoTME has 10.8M parameters and reports a favorable runtime–accuracy trade-off.

Critical weaknesses

My first concern is that the evidence for the broad multi-degradation restoration claim is somewhat indirect. The full-reference benchmark, STSVD, is primarily a smoky–clean dataset, whereas the more diverse degradations motivating PIDC, including blood-related veiling, reflection, blur, and low contrast, are mainly present in the real-world datasets. These datasets have no aligned clean targets and are evaluated after dataset-specific reference-free adaptation using no-reference IQA and temporal metrics. These results are useful, but they do not directly establish that the different degradation types are faithfully restored rather than simply producing outputs favored by the no-reference metrics. This is particularly relevant for medical video enhancement, where anatomical distortion or removal of clinically meaningful structures is more consequential than perceptual appearance alone. My second concern is that the temporal evaluation is not fully independent of the temporal machinery used by the method. EndoTME relies on optical flow for trajectory construction, includes an optical-flow-based temporal consistency loss during training/adaptation, and evaluates temporal quality using tWE and tOF with the same flow-estimation setup. This may favor methods whose outputs are explicitly optimized to be consistent under that estimator. An evaluation with an independent flow estimator or a complementary flow-free temporal metric would make the temporal-consistency claim more convincing.

Questions for authors

Q1: Are all methods in Table 2 adapted on the same LSVD and EndoReal splits with comparable training budgets? If not, please clarify the adaptation protocol for each baseline and provide a fair matched-adaptation comparison.

Q2: RTMR, the temporal loss, and the tWE/tOF metrics all rely on optical flow. Please evaluate temporal consistency with an independent flow estimator or a flow-free temporal metric to rule out possible metric coupling.

Q3: STSVD mainly provides paired smoky-clean data, while PIDC targets several additional degradations, including blood veiling, specular reflection, blur, and low contrast. Please provide degradation-specific quantitative results for these cases.

Q4: The real-world evaluation relies mainly on no-reference image-quality and temporal metrics. Can the authors provide an anatomical or downstream-task evaluation, such as segmentation, tracking, landmark detection, or clinician assessment, to verify that enhancement does not alter clinically relevant structures?

Q5: How much of PIDC’s gain comes from the specific analytical degradation cues rather than from simply adding extra spatial conditioning? Please include a control using learned degradation features or shuffled/mismatched cue maps while keeping the routing architecture unchanged.

4 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer CasN [part 1]

We appreciate your valuable feedback and support of our work.

For clarity, we refer to the reviewer's first concern regarding multi-degradation evaluation and anatomical fidelity as W1, and the second concern regarding the independence of temporal evaluation as W2.

To Q1:

All methods in Table 2 use identical data splits and strictly matched adaptation budgets within each target dataset.

Data splits. LSVD follows the official 416/70 training/test split, with the 416 training clips further divided into 374 adaptation and 42 validation clips, while the official 70 test clips remain strictly held out. EndoReal is partitioned at the patient level into adaptation, validation, and held-out test sets, comprising 256, 32, and 32 patients and 4,481, 598, and 612 video clips, respectively.

Adaptation and training budgets. All six methods undergo target-domain adaptation separately on LSVD and EndoReal under matched optimization budgets. Specifically, all methods receive 16k optimizer updates on LSVD and 84k on EndoReal, with the same effective number of target frames per update. Each method retains its original architecture and input formulation, ensuring identical target-domain data access and matched training opportunities within each dataset. Checkpoint selection and evaluation. All methods select checkpoints using the corresponding validation sets, while the held-out test sets are reserved exclusively for final evaluation. All methods are evaluated on identical test sequences using the same preprocessing, evaluation scripts, and metric implementations.

Therefore, Table 2 already presents a fair matched-adaptation comparison under controlled data splits, optimization budgets, and evaluation protocols.

To Q2 and W2:

We thank the reviewer for raising the concern about potential dependence on optical-flow estimation in temporal consistency evaluation. We would like to clarify that our original experiments already employed an independent optical-flow estimator for temporal consistency evaluation. Specifically, SpyNet is used for trajectory construction in RTMR and for computing the temporal consistency loss, whereas tWE and tOF are computed using an independent GMFlow estimator. GMFlow is used exclusively for evaluation and is independent of the SpyNet estimator employed within the model. All compared methods are evaluated using the same GMFlow estimator and evaluation protocol to ensure a consistent and fair comparison.

To Q3 and W1:

To address the concern regarding limited degradation-specific quantitative evidence, we conducted additional comparisons on four degradation-specific subsets of the paired STSVD test data, covering Blood Veil, Specular Reflection, Motion Blur, and Low Contrast. All compared methods are evaluated on the same test samples using a unified evaluation protocol, with PSNR and SSIM reported for each degradation category.

Table R1. Degradation-specific quantitative comparison. Each entry reports PSNR (dB) ↑ / SSIM ↑.

MethodBlood VeilSpecular ReflectionMotion BlurLow Contrast
RVRT22.38 / 0.848122.74 / 0.841521.17 / 0.893119.56 / 0.7096
DiffUIR24.42 / 0.894924.40 / 0.875125.53 / 0.921921.75 / 0.8165
SelfSVD24.15 / 0.879823.62 / 0.872524.08 / 0.886920.87 / 0.7983
MedVSR21.82 / 0.810920.59 / 0.809320.71 / 0.918619.34 / 0.7650
EndoIR24.87 / 0.880424.91 / 0.874825.09 / 0.902320.23 / 0.8177
EndoTME26.31 / 0.900825.97 / 0.890625.63 / 0.931221.46 / 0.8294

EndoTME achieves the highest SSIM across all four degradation categories and the best PSNR in three categories: Blood Veil, Specular Reflection, and Motion Blur. For Low Contrast, EndoTME achieves the second-best PSNR while maintaining the highest SSIM. These degradation-specific full-reference results provide additional quantitative support for the effectiveness of our framework under diverse visual degradations.We will incorporate these results and the corresponding evaluation protocol into the revised manuscript.

AuthorsOpenReview ↗

Rebuttal to Reviewer CasN [part 2]

To Q4 and W1:

To address the reviewer's concern regarding the preservation of clinically relevant anatomical structures, we conducted an additional downstream semantic segmentation experiment on CholecSeg8k to assess whether enhanced images retain important structural information in surgical scenes. CholecSeg8k contains 8,080 annotated frames from 17 laparoscopic videos, covering 13 semantic classes with pixel-level annotations. We adopt a video-level split of CholecSeg8k, with 9 videos (3,520 frames) for training, 3 videos (1,920 frames) for validation, and 5 videos (2,640 frames) for testing. We employ SegFormer-B2, initialized with ImageNet-1K pretrained weights, as a unified segmentation model. To fairly assess the impact of different enhancement methods on downstream segmentation, the segmentation model is trained exclusively on raw images and then applied to both raw test images and the corresponding images enhanced by each method, without method-specific retraining. All methods share the same segmentation model and ground-truth annotations within each test split. We report mIoU and mDice to evaluate segmentation performance.

Table R2. Downstream semantic segmentation results on CholecSeg8k.

MethodmIoU ↑mDice ↑
Raw0.50420.5725
RVRT0.67050.7831
DiffUIR0.68260.8057
SelfSVD0.50290.6372
MedVSR0.38070.5114
EndoIR0.68380.8046
EndoTME0.69370.8158

EndoTME achieves the highest mIoU (0.6937) and mDice (0.8158) among all compared methods, improving upon raw images by 0.1895 and 0.2433, respectively. Under the same segmentation model and annotation protocol, these results indicate that EndoTME-enhanced images better support downstream semantic segmentation under a fixed segmentation model, providing complementary evidence of structural information preservation.We have added the downstream segmentation results to Appendix G.3 (Line 1255) of the revised manuscript.

To Q5:

We thank the reviewer for the valuable suggestion to further examine the contribution of analytical degradation cues beyond additional spatial conditioning. To address this concern, we consider two complementary controls: the Learnable Tokens variant reported in Table 5 and a newly introduced Mismatched Cue Maps experiment. The former replaces degradation-derived conditioning with learnable tokens, while the latter replaces the six analytical cue maps entering PIDC with maps from different videos during inference. In the mismatched-cue experiment, RTMR, the spatial routing architecture, the three restoration experts, and all model parameters remain unchanged.

Table R3. Ablation of analytical degradation cues.

ConfigurationPSNR ↑SSIM ↑
Learnable Tokens24.710.8842
Mismatched Cue Maps23.490.8734
Full PIDC25.880.8922

Full PIDC outperforms Learnable Tokens by 1.17 dB in PSNR and 0.0080 in SSIM, demonstrating the benefit of degradation-derived conditioning over generic learnable tokens. Furthermore, mismatching the cue maps reduces PSNR by 2.39 dB and SSIM by 0.0188 compared with Full PIDC, despite using the same trained model and routing architecture. These results suggest that the correspondence between analytical cue maps and input image content is important for effective degradation-conditioned restoration.

We welcome further discussion if any concerns remain.

Reviewer_CasNOpenReview ↗

Response to author

Thank you for your response. I believe you have addressed most of my concerns; although a few questions still remain, I will give you a higher rating.

AuthorsOpenReview ↗

Response to CasN

Thank you for your recognition and support of our work and rebuttal. If you have any remaining questions or concerns, we would be happy to provide further clarification.

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

The paper targets endoscopic video enhancement and argues that temporal propagation is limited less by motion alignment than by whether the aligned history is worth using. EndoTME scores each trajectory-aligned historical candidate by appearance similarity, local quality, flow reliability and a learned degradation-risk term, keeps the top three per query, and aggregates them with a shifted state-space block. A second module, PIDC, derives six handcrafted cue maps (red veil, specular, blur, contrast, structure, flow), encodes them as type tokens and a dense spatial map, and uses them to condition features and to route three spatial experts. On the paired STSVD dataset it reports 25.88 dB PSNR and the best LPIPS, DISTS, tWE and tOF in the table, with no-reference gains on LSVD and on a private EndoReal collection.

Critical strengths
  1. The ablation design is the control most papers skip: Table 3 is a full 2x2 factorial, Tables 4-5 vary one module with the other held fixed, and the Learnable Tokens row of Table 5 replaces the physics cues with learnable tokens inside the same architecture (-1.17 dB). That row is precisely the capacity-versus-mechanism test I normally have to request, and it supports the claim that the cues rather than the extra parameters do the work.

  2. The trade-off is genuinely favourable: 10.80 M parameters and 27.36 ms per frame, only 0.61 ms slower than RVRT while gaining 3.85 dB PSNR, with every other video baseline at least twice as large. For an intraoperative setting this is worth more than a further 0.2 dB.

Critical weaknesses
  1. Only five baselines are used, and STANet — the method that introduced STSVD and the closest domain-specific competitor — appears only in prose, so the state-of-the-art claim on that dataset is not tested against the model it was built for. The stated reason (no released code or checkpoints) does not prevent tabulating its published numbers on the same split, and PromptIR and AirNet, the two degradation-conditioning methods PIDC claims to extend, are never compared at all.

  2. All full-reference evidence comes from STSVD, a synthetic smoke dataset, while four of the six cues target degradations — specular, blur, contrast, motion — that are never evaluated with a reference. Real-world evidence is no-reference on a private 5,691-clip set whose EndoReal margins are thin (BRISQUE 40.94 vs 41.45, CLIP-IQA 0.5511 vs 0.5469) and whose checkpoint-selection signal is never stated.

Questions for authors

Q1. Can you tabulate STANet's published numbers on the STSVD split, and add PromptIR or AirNet as degradation-conditioning baselines? Without them the increment over the line PIDC claims to extend is unmeasured.

Q2. The degradation-risk term contributes 0.86 dB of RTMR's 1.53 dB, more than the three handcrafted cues combined (0.67 dB). Is that from reliability-aware selection or from the learned gate — what does a shuffled or frozen degradation-risk score give?

Q3. What signal selected the LSVD and EndoReal checkpoints, given that neither has a clean reference? If a no-reference metric was used, the reported NIQE, BRISQUE and CLIP-IQA are partly selected for.

Q4. Only synthetic smoke supports full-reference conclusions, yet the cues target five degradation types. Is there a per-degradation breakdown — a specular or motion-blur subset, or an ablation dropping one cue — showing which cues actually fire and help?

Q5. What are the parameter counts of each row in Tables 3-5, in particular Learnable Tokens versus Full PIDC? Please confirm that the -1.17 dB holds at matched capacity.

4 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer oair [part 1]

We appreciate your valuable feedback and support of our work.

To Q1 and W1:

We thank the reviewer for the valuable suggestion regarding baseline coverage and comparison fairness. Following this suggestion, we have extended the quantitative comparisons on STSVD by including three additional representative restoration methods: STANet, PromptIR, and AirNet. Among them, PromptIR and AirNet are representative degradation-conditioning approaches. All additional experiments follow the same data splits and evaluation protocol to ensure fair comparisons.

Table R1. Additional quantitative comparison on STSVD.

MethodPSNR (dB) ↑SSIM ↑LPIPS ↓DISTS ↓tWE ↓tOF ↓
STANet23.640.88260.23790.17940.15730.2452
PromptIR24.950.88960.14250.08610.14860.1967
AirNet21.230.87040.09810.10060.13620.2781
EndoTME (Ours)25.880.89220.08400.07870.12870.1736

The results demonstrate that EndoTME achieves the best restoration performance under the unified experimental protocol, further validating the effectiveness of our approach for endoscopic video restoration. We have added these comparisons to Appendix G.2 (Line 1236) of the revised manuscript.

To Q2:

We thank the reviewer for the insightful question regarding the contribution of reliability-aware selection and the learned gate. To further validate the effectiveness of degradation-risk estimation, we introduce a Shuffled Risk control, which randomly permutes the predicted risk scores among valid historical candidates while keeping the model parameters, other reliability cues, and residual gate unchanged.

Table R2. Shuffled-risk control on STSVD.

ConfigurationPSNR (dB) ↑SSIM ↑
w/o Degradation-Risk25.020.8865
Shuffled Risk25.370.8843
Full RTMR25.880.8922

Compared with Shuffled Risk, Full RTMR achieves 0.51 dB higher PSNR and 0.0079 higher SSIM. These improvements demonstrate the importance of correctly associating degradation-risk scores with historical candidates for effective memory retrieval and aggregation. Since the residual gate and all other model parameters remain unchanged, this comparison provides evidence that the candidate-specific association of degradation-risk scores contributes to reliable historical feature selection and aggregation.

To Q3 and W2:

We thank the reviewer for the question regarding checkpoint selection. For LSVD and EndoReal, we use the reference-free adaptation loss defined in Appendix D.2 (Eq. 48), which combines temporal consistency, structural preservation, and residual regularization. Both datasets are independently adapted from the STSVD-pretrained checkpoint. During adaptation, we evaluate the weighted adaptation loss on the corresponding validation split and select the checkpoint with the lowest total validation loss for final testing. NIQE, BRISQUE, and CLIP-IQA are computed only for final evaluation, with the held-out test sets excluded from checkpoint selection. Real-world evaluation covers both the public LSVD dataset and EndoReal. As shown in Table 2, EndoTME achieves the best results across all five metrics on both datasets. On EndoReal, it achieves NIQE of 6.42 (vs. 7.67), tWE of 0.1041 (vs. 0.1117), and tOF of 0.1015 (vs. 0.1136), further supporting its real-world restoration quality and temporal consistency.

AuthorsOpenReview ↗

Rebuttal to Reviewer oair [part 2]

To Q4 and W2:

We thank the reviewer for the valuable suggestion regarding degradation-specific quantitative evaluation. Following this suggestion, we have conducted additional quantitative comparisons on the paired STSVD data, covering four degradation conditions: Blood Veil, Specular Reflection, Motion Blur, and Low Contrast. All compared methods are evaluated on the same test samples using a unified evaluation protocol, with PSNR and SSIM reported for each degradation category.

Table R3. Degradation-specific quantitative comparison. Each cell reports PSNR (dB) ↑ / SSIM ↑.

MethodBlood VeilSpecular ReflectionMotion BlurLow Contrast
RVRT22.38 / 0.848122.74 / 0.841521.17 / 0.893119.56 / 0.7096
DiffUIR24.42 / 0.894924.40 / 0.875125.53 / 0.921921.75 / 0.8165
SelfSVD24.15 / 0.879823.62 / 0.872524.08 / 0.886920.87 / 0.7983
MedVSR21.82 / 0.810920.59 / 0.809320.71 / 0.918619.34 / 0.7650
EndoIR24.87 / 0.880424.91 / 0.874825.09 / 0.902320.23 / 0.8177
EndoTME26.31 / 0.900825.97 / 0.890625.63 / 0.931221.46 / 0.8294

EndoTME achieves the highest SSIM across all four degradation-specific subsets and the best PSNR in three subsets: Blood Veil, Specular Reflection, and Motion Blur. On the Low Contrast subset, it achieves the second-best PSNR while maintaining the highest SSIM, further supporting its effectiveness under diverse visual degradations.

To Q5:

We thank the reviewer for the suggestion to clarify the parameter counts and verify the capacity-matched comparison. We report the trainable parameter counts for all configurations in Tables 3–5, following the same counting convention as the complexity analysis in the main paper.

Table R4. Trainable parameter counts of ablation configurations in Tables 3–5.

Ablation groupConfigurationPSNR (dB) ↑Trainable params (M)
Module-levelw/o RTMR & PIDC23.129.976550
Module-levelRTMR only24.3410.081322
Module-levelPIDC only24.2510.699467
Module-levelFull Model25.8810.804239
RTMRTrajectory Alignment Only24.3510.803436
RTMR+ Feature Similarity24.7110.803436
RTMR+ Local Quality24.6310.803533
RTMR+ Flow Reliability25.0210.803598
RTMR+ Degradation Risk25.8810.804239
PIDCLearnable Tokens24.7110.804239
PIDCw/o Type-Level Tokens24.9710.609711
PIDCw/o Dense Spatial Embedding25.4610.799695
PIDCFull PIDC25.8810.804239

The Learnable Tokens variant preserves the same number, dimensionality, and trainable parameter count of the tokens, but does not modulate them using input-dependent analytical degradation cues.These results confirm that the 1.17 dB improvement is obtained at matched trainable parameter capacity, supporting the effectiveness of analytical degradation cue-based conditioning beyond simply increasing the number of learnable parameters.

We welcome further discussion if any concerns remain.

After considering the responses in the rebuttal and the other reviews, I retain my original rating of the paper.

AuthorsOpenReview ↗

Response to Reviewer oair

We sincerely appreciate your thoughtful consideration of our rebuttal and your continued support of our work.

Reviewer_BW9VOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Prior methods primarily focus on motion alignment and temporal fusion of historical information, without explicitly assessing whether aligned historical observations are degraded and worth using. This work distinguishes motion correspondence from information reliability and combines local quality, motion stability, and degradation cues to select reliable historical evidence for restoration.

Critical strengths

A clear insight linking temporal reliability and degradation-aware restoration. The paper distinguishes accurate motion correspondence from useful historical evidence: even correctly aligned observations can propagate degradation. RTMR addresses this through selective retrieval, while PIDC adapts restoration to local degradation cues. This gives the framework a coherent motivation and addresses a meaningful limitation in endoscopic video restoration.

Critical weaknesses

The description does not clearly distinguish degradation identification from heuristic cue extraction. The main text suggests that PIDC characterizes different endoscopic degradations, whereas Appendix C.1 defines analytical proxies based on color, contrast, gradients, and motion. It remains unclear how these cues distinguish degradation from normal tissue appearance—for example, blood-related veiling from naturally reddish, low-texture tissue.

Questions for authors

Is PIDC intended to identify specific degradation types, or to provide heuristic visual cues for restoration? In particular, how does it avoid interpreting naturally reddish, low-texture tissue as blood-related veiling?

3 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer BW9V

We appreciate your valuable feedback and support of our work.

To Weakness and Question:

We thank the reviewer for highlighting the ambiguity in the description of PIDC. We clarify that PIDC employs physics-inspired visual cues to provide degradation-conditioned information for restoration. As detailed in Appendix C.1, PIDC extracts six analytical cues: Red-veil, Specular Reflection, Local Blur, Weak Contrast, Structural Response, and Flow Magnitude. These cues characterize local degradation-related visual conditions through color, intensity, local statistics, gradients, and motion information.

Regarding naturally reddish, low-texture tissue, we acknowledge that an individual Red-veil response can be visually ambiguous, as normal reddish tissue with limited texture may also produce a high response. PIDC reduces its reliance on any single cue by incorporating complementary visual cues and image-content features for adaptive restoration. Specifically, the Red-veil cue is processed alongside structural response, local contrast, and motion information to construct type-level tokens and dense spatial embeddings. The type-level tokens interact with RTMR trajectory features through cross-attention, allowing the conditioned representation to incorporate both degradation-related cues and historical visual information. The spatial router further combines these features to predict location-dependent soft weights over three restoration experts, adapting restoration behavior according to local image content and multiple visual cues. Thus, even when normal tissue produces a high Red-veil response, this response does not directly trigger a predefined blood-removal operation. The model can incorporate additional visual evidence to modulate restoration without relying on the Red-veil response alone.

We have clarified the description of PIDC in the revised manuscript (Section 3.3, Line 226) to more accurately explain how analytical degradation cues guide spatially adaptive restoration.

We welcome further discussion if any concerns remain.

The author's rebuttal solved my question, I will raise my score.

AuthorsOpenReview ↗

Response to Reviewer BW9V

Thank you for your positive feedback and for recognizing our work. We sincerely appreciate your time and support.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Responses to Reviewers

We sincerely thank all reviewers for their careful evaluation, valuable comments, and constructive suggestions.

We have provided detailed, point-by-point responses to the reviewers' questions and concerns in our rebuttal. We have also made targeted revisions to the manuscript to address the key issues raised during the review process. The main changes in the revised PDF are summarized below:

  1. PIDC Clarification (Section 3.3, Line 226): Refined the description of PIDC to clarify how analytical degradation cues guide spatially adaptive restoration.

  2. Additional Baseline Comparisons (Appendix G.2, Line 1236): Added quantitative comparisons with STANet, PromptIR, and AirNet on STSVD.

  3. Downstream Semantic Segmentation (Appendix G.3, Line 1255): Added downstream segmentation results on CholecSeg8k to further assess the preservation of task-relevant structural information.

We sincerely appreciate the reviewers' time and thoughtful feedback, which have helped us improve the clarity and experimental evaluation of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.