List of Papers Browse by Subject Areas Author List
Abstract
Super-resolution (SR) models are attracting growing interest for enhancing minimally invasive surgery and diagnostic videos under hardware constraints. However, valid concerns remain regarding the introduction of hallucinated structures and amplified noise, limiting their reliability in safety-critical settings. We propose a direct and practical framework to make SR systems more trustworthy by identifying where reconstructions are likely to fail. Our approach integrates a lightweight error-prediction network that operates on intermediate representations to estimate pixel-wise reconstruction error. The module is computationally efficient and low-latency, making it suitable for real-time deployment. We convert these predictions into operational failure decisions by constructing Conformal Failure Masks (CFM), which localize regions where the SR output should not be trusted. Built on conformal risk control principles, our method provides theoretical guarantees for controlling both the tolerated error limit and the miscoverage in detected failures. We evaluate our approach on image and video SR, demonstrating its effectiveness in detecting unreliable reconstructions in endoscopic and robotic surgery settings. To our knowledge, this is the first study to provide a model-agnostic, theoretically grounded approach to improving the safety of real-time endoscopic image SR. Code is available: https://github.com/jusiro/Endoscopic-CFM
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3256_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/jusiro/Endoscopic-CFM
Link to the Dataset(s)
N/A
BibTex
@InProceedings{SilJul_Trustworthy_MICCAI2026,
author = { Silva-Rodríguez, Julio AND Konukoglu, Ender},
title = { { Trustworthy Endoscopic Super-Resolution } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16882},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a novel framework for creating “Conformal Failure Masks” (CFMs) for endoscopic super-resolution (SR) models. The goal is to identify and flag unreliable regions in the SR output, thereby enhancing trust and safety in clinical applications. The method involves two main stages: first, training a lightweight, auxiliary “Reconstruction Error Network” on the intermediate features of a pre-trained SR model to predict pixel-wise reconstruction error. Second, it applies the principles of conformal risk control to convert these continuous error predictions into a binary failure mask. This process provides a theoretical, distribution-free guarantee that the false-negative rate (i. e. , the rate of undetected failures) will be below a user-specified level. The approach is evaluated on endoscopic and robotic surgery SR tasks, demonstrating its ability to control risk and improve the effective image quality in trusted regions.
- 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.
This is a well-written paper with several significant strengths that make it a valuable contribution.
-
High Clinical Relevance and Novel Application: The paper addresses a critical and practical problem: ensuring the trustworthiness of generative models like SR in safety-critical medical applications. The risk of “hallucinated” details in SR-enhanced endoscopy is a major barrier to clinical adoption. The proposed CFMs offer a clear and actionable clinical utility by providing an explicit “do-not-trust” signal to the practitioner, which could support more informed decision-making and improve patient safety. To my knowledge, this is one of the first works to propose a theoretically grounded safety-masking framework specifically for endoscopic SR.
-
Methodological Soundness and Clarity: The approach is methodologically sound, clearly explained, and well-motivated. By decoupling the error prediction from the SR model itself, the framework is model-agnostic and efficient, as it does not require re-training the base SR model. The use of conformal risk control is a major strength, as it provides a rigorous, theoretically grounded basis for the failure masks. This is a significant step up from more heuristic uncertainty measures that lack formal guarantees.
-
Empirical Results: The experimental results presented in Table 1 are compelling. The method successfully controls the empirical false-negative rate (FNR) close to the desired level alpha, validating the theoretical promise of the conformal procedure. Furthermore, the analysis shows that by excluding the regions flagged by the CFM, the average PSNR of the remaining “trusted” parts of the image increases significantly. This demonstrates that the framework is not just theoretically sound but also practically effective.
-
- 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.
-
Literature Positioning: The paper’s claim in the introduction that Conformal Prediction (CP) is “relatively underexplored in medical image computing” is an overstatement. In recent years, there has been a significant surge of CP research in the medical imaging community, including at top venues like MICCAI, with applications spanning classification, segmentation, and regression. The authors should tone down this claim and provide a more accurate and nuanced positioning of their work within this growing body of literature. Similarly, while the paper does compare to the recent work of Adame et al. , a more detailed and upfront discussion in the introduction or related work section would better highlight the specific contributions and differences of the proposed bi-level risk control from the outset.
-
Unsubstantiated Claims of Real-Time Performance: The abstract and introduction repeatedly motivate the work by its suitability for “real-time” applications, which is a key selling point for a clinical tool. However, the paper provides no rigorous evaluation of this claim. While a single latency number is mentioned, there is no comprehensive analysis of the added computational overhead of the Reconstruction Error Network and the conformalization step. To substantiate the real-time claim, a proper evaluation of inference time and throughput, comparing the base SR model with the full CFM pipeline, is necessary.
- Figure and Caption Quality: The figures, while containing useful information, could be significantly improved.
- Figure 1 is the most prominent example. It lacks a descriptive caption explaining what is being shown in the various insets (e. g. , A, B, C, G). The reader has to hunt for the explanation on page 8, which disrupts the flow. A self-contained caption is essential.
- Figures 2 and 3 are dense. The overall presentation feels squished.
- Visualization of CFMs: The qualitative examples of CFMs in Figure 1 sometimes appear as noisy, scattered masks which might be difficult to interpret clinically. For better practical utility, it might be beneficial to visualize the masks after applying some form of morphological processing (e. g. , removing small, isolated regions) or by displaying the underlying error prediction map thresholded to show a more coherent “failure region. “
(The link to the anonymous repository did not work for me)
-
- Please rate the clarity and organization of this paper
Satisfactory
- 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 am recommending “Weak Accept” because the paper tackles an important and relevant clinical problem with a method that is novel, elegant, and theoretically sound. The core idea of creating Conformal Failure Masks for SR is a valuable contribution, and the results convincingly demonstrate its effectiveness. The paper’s strengths, particularly its clear clinical utility and methodological rigor, are significant.
However, the paper is held back by several notable weaknesses that prevent a stronger recommendation. The exaggerated claims about the novelty of CP in the field, the unsubstantiated real-time performance claims, and the poor quality of the figures and referencing need to be addressed in a final version. I believe the paper is promising and can become a strong contribution, but it requires a careful revision to correct these shortcomings.
- 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 concerns raised have been mostly addressed and since I believe the methodological contribution is valuable, I opt for accepting the paper.
Review #2
- Please describe the contribution of the paper
The paper engages with the problem of untrustworthiness of super-resolution networks. Super-resolution upscales images, which can generate invalid pixels. This work generates masks to indicate pixels that are likely to have errors. They additionally introduce an configurable alpha parameter that can set False Positive Rate (FPR). Configuring this parameter allows the FPR of the masks to be set as desired by the user. Their experiments show close alignment of the desired alpha and the actual FPR value. Their experiments also show their masks identify areas with errors, as the PSNR is higher on the non-masked (trustworthy) areas.
- 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.
- a method based on the intermediate features of the super-resolution (SR) network, rather than the confidence outputs of the SR network.
- A configurable FPR value
- Improved PSNR scores for pixels marked as trustworthy
The paper clearly sets out the technical challenge and the method used. The addition of the configurable alpha parameter is novel. The alpha parameter does not appear in previous literature, to my knowledge. The alpha parameter has clear applicability and the paper’s analysis of the FPR shows a strong correlation with the desired alpha. This engages not only with the current problem but begins to engage with future challenges that an application would need to address. Their results support their claims that their approach engages with the problem set out.
- 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.
- It is not clear how this would be applied practically. Identifying untrustworthy pixels does not resolve the issue that those pixels cause. Please suggest how, or reference literature that shows how these masks can be used for positive impact to a surgeon. E.g. by applying these masks in real time, a surgeon would know to be cautious around areas with the mask….
- The literature review mentions that other techniques use the confidence values to perform a similar task, yet there is no comparison to other techniques. It is therefore not clear how their method performs compared to other techniques that would mask untrustworthy outputs. Please add a comparison to those techniques (or another appropriate benchmark) to compare your novel technique to existing methods
- Although the technique uses intermediate features rather than output confidence scores, it is not clear why this is superior to the confidence scores. Please provide more justification for this. This will be supported by eviencing the comparison to existing techniques.
- Please add some discussion on the limitations of the technique when the super-resolution model is performing poorly. When the super resolution model is performing very poorly overall, it is likely that the input data is outside of the training data distribution. In this scenario, the CFM would be most needed, but also may perform poorly. Justification of this or a discussion of this limitation or data to disprove this concern would strengthen the paper.
- 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
Fig 1:
- is referenced in the introduction, but the caption and introduction are insufficient to understand it. The explanation provided at the end of section 3 should be moved to the caption of the image to help understand it.
- Diagram text is black on a dark background, please make it lighter to be clearer
- Sub-fig B appears to repeat the output of the super resolution rather than showing the high resolution original.
- Please consider adding the masks to the sub-figure outputs to make it clear where the masks are. Table1:
- It is not clear why you have included bicubic, a comparison of bicubic vs SwinIR is not of interest, the paper is about the CFM.
- As mentioned in the weaknesses, please add a comparable technique. In particular a method that uses the confidence scores rather than the intermediate features.
Equation 1:
- It seems this is the loss function of the Reconstructive Error network? Please make that clear. I.e. add an L = to show it is the loss function
I found the description of the bi-level optimisation difficult to follow. The first 3 lines appear to repeat the same point as the last 3 lines of the previous subsection. I may have misunderstood. Some clarifying language would help here.
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Currently the novelty and the practical applicability is too low to accept. With recommendations applied, this could improve.
The major reasons for my recommendation are:
- The lack of comparison with the existing methods that use the confidence scores
- The lack of clear application of these masks
- Unclear that CFM performs acceptably when the super-resolution model performs poorly
The alpha parameter is a clear practical strength of the paper and use of intermediate features is somewhat novel but not very novel and lacks supporting justification. These strengths do not outweigh the weaknesses identified.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The main contribution of the paper is the introduction of a model-agnostic framework for trustworthy super-resolution in medical imaging, which combines: (1) a lightweight reconstruction error prediction network, and (2) a bi-level conformal risk control procedure to generate Conformal Failure Masks.
- 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) The introduction of Conformal Failure Masks, which offer interpretable and operational outputs. 2) The lightweight design of the Reconstruction Error Network.
- 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) The methodological novelty is limited, as the proposed framework largely combines existing techniques, namely auxiliary error prediction networks and conformal prediction, without introducing fundamentally new theoretical developments. 2) The conformal guarantees rely on standard assumptions such as exchangeability, which are unlikely to hold in real-world medical imaging settings with domain shifts and structured spatial dependencies. 3) The evaluation lacks strong clinical validation, as experiments are limited to benchmark datasets without involving clinicians or demonstrating impact on downstream decision-making. 4) The comparison with prior work is limited. 5) The error prediction network is relatively simple and may limit performance, yet no ablation study is provided to justify its design choices or explore more advanced alternatives.
- Please rate the clarity and organization of this paper
Satisfactory
- 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
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 addresses an important problem in trustworthy medical image super-resolution and proposes a practical and interpretable solution. However, the overall contribution is limited by the lack of strong methodological novelty, incomplete comparison with prior work, and insufficient clinical validation.
- 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’ response has addressed some of my previous concerns. Therefore, between accept and reject, I am inclined to choose accept.
Author Feedback
We thank the reviewers for their time, positive comments, and feedback for improvement.
1.Comparison to prior work (R2,R3). Our goal is operational, user-controlled failure detection for pretrained real-time image/video SR models, while remaining SR-model agnostic and lightweight.
- Baselines. SR is a regression task: standard SR models output continuous images, not confidence scores as in discriminative calibration. Thus, confidence-score baselines are not directly available without modifying/retraining the SR model. Ensembles, MC dropout, and Bayesian variants require multiple forward passes and/or stochastic layers, which add latency and break compatibility with common SR architectures. Moreover, these methods alone do not yield operational binary failure decisions with distribution-free control of missed failures. Therefore, existing uncertainty methods are not direct baselines under our constraints; this gap highlights the problem’s complexity and the novelty of our procedure.
2.R1’s comments.
- Positioning. We will correct overclaims and narrow the CP novelty scope to SR/generative tasks. We will discuss Adame’s work earlier.
- Visualizations & figures. We agree that figures/captions were condensed due to space constraints and we will add the suggestions.
- Real-time. SwinIR-2×/BasicVSR run at ~30/20 ms per image (~33/50 FPS). The error network has ~70K parameters and adds ~5 ms; thresholds are precomputed, and masks require only thresholding. Full pipeline remains ~28/45 FPS. We will report base-SR vs full-pipeline latency.
3.R2’s comments.
- Confidence scores & intermediate features. SR models do not provide confidence scores natively, since they output continuous regression responses rather than discriminative scores. Intermediate features before the decoder are common to image/video SR models, motivating their use to learn our error measure.
- Application. CFMs are safety overlays, not corrections of failed pixels. Masked regions indicate where SR should not be trusted: users may inspect the raw image, reduce reliance on SR, or spend more time assessing suspicious/noisy structures. We will clarify these use cases.
- Poor SR performance. Since SR pre-training and failure detection are decoupled, our conformal procedure maintains guarantees under exchangeability. Eq.2 controls the error level (PSNR with tau_fail). Thus, if SR performs poorly but must be used, the practitioner can set a lower PSNR limit; if a high PSNR is required with a weak SR model, CFMs yield larger rejected regions, which is the desired conservative behavior.
4.R3’s comments.
- Clinical validation. We agree this is valuable future work. Our current goal is to establish the methodological basis and quantitative, reproducible validation of this paradigm using public datasets/models, and to foster community discussion at MICCAI on how to practically use and assess these masks.
- Novelty. We introduce novel operational masks with guarantees to locate hallucinated outputs in real-time SR. Technically, the bi-level conformal procedure (Eq.2) is novel, giving users control with guarantees on the reconstruction error level and FNR. This enables extending auxiliary error prediction networks to regression problems, where there is no intrinsic notion of correctness, unlike their prior discriminative use.
- Error network simplicity. It is intentionally simple to preserve real-time latency and model agnosticism. Heavier alternatives, e.g., attention, add overhead at large resolutions. Its design is supported by its strong error detection (Fig.2; AUROC >98% across 3 datasets), and the complexity ablation in Fig.4 did not improve results.
- Guarantees. Exchangeability assumptions were acknowledged. Measures such as weighted conformal prediction could address violations, but warrant future dedicated studies. Empirically, HyperKvasir spans diverse gastro/colonoscopy cases, and we observed FNR close to the target level on fresh held-out endoscopy procedures.
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.
Please prioritize addressing concerns regarding novelty and positioning with respect to prior work, including comparisons to existing confidence-based approaches and clearer justification of design choices (e.g., the use of intermediate features and the error prediction network). Additionally, clarify the practical utility of the method and its reliability when the base model performs poorly. Also, address concerns regarding the validity of the conformal guarantees.
- 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 presents a model-agnostic framework for trustworthy super-resolution in medical imaging, combining a lightweight reconstruction-error prediction network with a bi-level conformal risk control procedure to generate Conformal Failure Masks. The work is clinically relevant and presents a novel application of conformal methods to an important reliability problem in medical image reconstruction. The empirical results are strong and support the effectiveness of the proposed approach. A significant concern raised during review was the limited comparison with existing work and the need to better position the contribution relative to prior literature. The authors’ rebuttal provides an acceptable response to this issue, and I expect the final revision to further clarify the paper’s positioning to existing methods. Overall, the remaining concerns are not severe enough to outweigh the clinical relevance, novelty, and empirical strength of the work. The paper is therefore recommended for acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The authors did not manage to convincingly reply to the reviewers comments and some suggestions have been left unsolved.
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 two reviewers that interacted in the post-rebuttal phase both mention that their concerns have been (mostly) addressed and that they now propose to accept the paper. Given that the one reviewer who did not enter a post-rebuttal decision initially rated the paper with a weak reject, this sways the paper to the positive side of the decision threshold. Moreover, having read both he original concerns of this reviewer and the author rebuttal, it seems that at least some of these concerns have also been properly addressed in the rebuttal.
