List of Papers Browse by Subject Areas Author List
Abstract
Critical retained foreign objects (RFOs) on intraoperative
Chest radiographs are rare but high-risk events, and their scarcity limits robust automated detection model training and generalization. We introduce SurgRFO, a two-stage synthesis framework for generating realistic RFO-present intraoperative chest X-rays. In Stage 1, a Roentgen chest X-ray foundation model is fine-tuned on surgical-domain images to generate realistic, RFO-free backgrounds that preserve anatomy, indwelling lines and tubes, and intraoperative imaging characteristics. In Stage 2, a lightweight generator trained on localized RFO patches from limited positive cases synthesizes diverse RFO instances, which are composited onto backgrounds using conditional Poisson fusion for photometric consistency. We evaluate SurgRFO through (i) a blinded clinician study assessing realism and clinical plausibility, and (ii) downstream detection performances when augmenting Faster R-CNN, YOLOv8, and Retina with synthesized data. SurgRFO consistently improves sensitivity at low false-positive-per-image (FPPI) operating points on internal and external test sets. Clinician ratings indicate realism comparable to real intraoperative images. Ablation analyses examine fusion strategies and synthesis scale, and ethical safeguards for synthetic surgical data are discussed. The code is publicly available on GitHub at https://github.com/Yuanyun936/SurgRFO/.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1483_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Yuanyun936/SurgRFO
Link to the Dataset(s)
N/A
BibTex
@InProceedings{HuYua_SurgRFO_MICCAI2026,
author = { Hu, Yuanyun AND Wang, Yuli AND Acevedo Rodriguez, Noemi AND Yang, Ronald AND Hsu, Wen-Chi AND Luo, Siwei AND Bai, Zihao AND Wu, Jing AND Dai, Yuwei AND Wu, Shaoju AND Lindquist, Jonathon AND Honce, Justin AND Trivedi, Premal AND Jiao, Zhicheng AND Kamel, Ihab AND Haut, Elliott AND Johnson, Pamela AND Eng, John AND Lin, Cheng Ting AND Su, Nan AND Chen, Bo AND Yu, Sun AND Bai, Harrison},
title = { { SurgRFO: Foundation Model Based Compositional Synthesis of Critical Retained Foreign Objects in Intraoperative Chest X-Rays } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16893},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose SurgRFO, which synthesizes intraoperative chest X-rays with “critical” retained foreign objects (RFOs) to improve downstream detection of these RFOs. The framework comprises two stages: first, a fine-tuned foundation model that generates chest radiographs, and second, a model that adds critical RFOs to the images through a blending mechanism. The authors claim they will release a benchmark dataset of 144 annotated intraoperative chest radiographs with critical RFOs (although this claim is refuted in the GitHub).
- 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.
- Significance of clinical application. Detecting RFOs is an important clinical problem (https://www. cureus. com/articles/320644-retained-foreign-object-signals-a-dangerous-atmosphere-in-the-operating-room#!/; https://www. sciencedirect. com/science/article/pii/S0039606025003654). While various safeguards have been explored to facilitate the detection of RFOs, RFOs continue to go undetected due to the complexity of the task.
- Use and release of clinical radiographs with critical RFOs. Due to the rarity of these objects, evaluating on these images and sharing them with the public is a significant contribution.
- Robust evaluation of synthetic image quality, including a blinded clinician study, quantitative distribution measures, and a downstream classification performance analysis across three classification models.
- External validation was performed.
- The code was made publicly available.
- The first figure is an informative graphical abstract.
- The proposed framework splits synthesis into two parts (radiograph then RFO) to strengthen the fidelity of the RFOs. The synthetic images improve downstream classification across three classification models
- 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.
- Lack of downstream performance comparison with other synthetic generation models/augmentation strategies. The work does not compare its downstream detection results against closely related prior methods, such as https://link. springer. com/chapter/10.1007/978-3-032-08009-7_15 that similarly applies synthetic data augmentation for detecting critical RFOs. In addition, the paper does not provide downstream performance comparisons against alternative generative models or augmentation strategies. As a result, it remains unclear whether the observed gains are attributable to the specific generative methodology proposed, or simply to increasing the effective training set size. Would we see similar performance gains with RoentGen? What about if we just injected RFO patterns into otherwise negative images?
- Substantial overlap with earlier work, limiting novelty. The work does not consider a 2025 MICCAI workshop paper (https://link. springer. com/chapter/10.1007/978-3-032-08009-7_15), which similarly employs synthetic data augmentation to improve downstream critical RFO detection and releases a synthetically generated dataset of chest radiographs with critical RFOs. The submission should have referred to and contextualized this work (i. e. compared model performance), clearly distinguishing the submission’s contributions from this prior work. Especially when making claims such as “no public dataset dedicated to critical RFOs exists” and “the first publicly available dataset dedicated to critical RFOs. ” The two contributions of using synthetic data augmentation to improve downstream critical RFO detection and the release of a chest radiograph dataset of critical RFOs were again repeated in an arXiv preprint (https://arxiv. org/abs/2507.06937v1) that the submission cited as a work with “high-fidelity and diverse X-ray synthesis”. While the preprint cannot limit the novelty of the submission under consideration as it is not published, it should at least be properly contextualized if cited.
- Conflicting information on the dataset release. The authors state in the manuscript that they are publicly releasing their dataset (no link is provided), but on GitHub, they state that they cannot release the data due to privacy constraints. As this dataset release is arguably one of the greatest contributions of the work, the GitHub message is concerning.
- Missing dataset details that are important for contextualizing the results and public release. How many patients were there? With a 6:1:3 patient-level split, how many radiographs did that put in the test dataset? What were the class ratios in the test dataset? It is important for the reader to know the number of scans in the test dataset, how many contained critical RFOs, and how many patients those scans were sourced from. In addition, how many radiologists labeled the images? Were multiple perspectives considered during this labeling process? If so, how were disagreements between radiologist labels managed, especially when it comes to bounding boxes? What were the criteria for differentiating between critical and non-critical RFOs? Was it truly only the training images that were preprocessed? Shouldn’t all images go through the same preprocessing pipeline? Was the same training dataset used for the synthesis network and the classification network? If so, for synthesis network training, I assume all the negatives in the training dataset constituted D_bg and the positives D_rfo? How was Stage-1 synthesis evaluated on 5k images when there were only 944+144 in the dataset? What radiographs were the external cohort compared to, as all reported external images had critical RFOs?
- Downstream detection performance measures were not reported with measures of uncertainty or confidence. Additionally, the authors did not perform a statistical analysis of the downstream performance gains. Both of which reviewers were supposed to look for in the official MICCAI guidelines (https://conferences. miccai. org/2026/en/REVIEWER-GUIDELINES. html). This is especially concerning as the reader does not know how small of a sample size is being tested, nor the class balance of that sample. Having said that, the proposed model has obvious enough performance gains over the baseline.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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 abstract says, “ethical safeguards for synthetic surgical data are discussed”, but this is not actually an aspect of the work.
- The introduction contains duplicate sentences: “Two fundamental challenges dominate. Two major challenges arise. “
- In the introduction, “low hundreds of images” doesn’t make much sense.
- Figure 1 introduces inconsistencies that can easily be misinterpreted. For example, it implies that data was sourced from three institutions, even though it was sourced from two. Furthermore, it implies radiologists looked through >50,000 images to locate RFOs, when really this is a retrospective study where the authors used reports to identify RFOs. The actual 5-point Likert scale would be more useful to see than the specified “Image Rating Rubric”. Was the question truly, “Which one is real?” Because the metric “Syn->Real” implies the question was “Is this image synthetic or real”.
- In 2.1, “Patient-level splits into training, validation, and test sets were applied consistently across cohorts to prevent information leakage” is redundant with 3, which makes much more sense “split into training, validation, and test sets with a 6:1:3 patient-level ratio. ”
- In 3.1, I would take a moment to discuss how adapted Roent-Gen worsened diversity performance (MS-SSIM) and why that’s okay.
- In Section 3.1, when discussing Table 2, the justification for treating FID as the primary metric and MS‑SSIM as secondary on the grounds that MS‑SSIM would be dominated by preserved backgrounds is unconvincing. FID is similarly sensitive to large‑scale structural similarity and would also be dominated by background preservation in this setting. As such, this distinction does not meaningfully address the underlying evaluation issue. It would be preferable to state the results directly and acknowledge the observed performance differences, rather than providing a rationale that does not apply consistently across metrics.
- In Table 3, I would bold both 0.394 for YOLOv81 because bolding one while underlining the other seems arbitrary
- The radiologist-driven Turing test would make more sense before the downstream evaluation, as it is evaluating synthetic image quality, which is directly related to section 3.1.
- Briefly describing the training of classification models would be helpful. Was hyperparameter tuning performed on the validation dataset for each model?
- I know that MICCAI has limited space, but seeing the actual results for external validation would be helpful.
- How was confidence measured for ConfAUC? How many images were in Turing test?
- 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?
I would have recommended acceptance this paper, if not for Weakness 1.As the main purpose of this paper was to improve downstream classification performance, I would have liked to see the paper prove to me that the proposed methodology outperformed other forms of augmentation. I really loved the idea of a two-step framework, but the results didn’t demonstrate to me that it was necessary. All the results really say is that synthetic augmentation improves classification performance, which is widely well-known.
While I would not like to see these additional experiments in a rebuttal, I would like to see that the authors appropriately contextualize past works, explicitly stating the contributions of their paper over the past works. I would also like to see confirmation that the radiological images with critical RFOs will be released (which would be the largest contribution of the work). And finally, I would want to see that the missing information has been added in, and that the erroneous sentence in the abstract claiming ablation analyses and ethical considerations was deleted.
- 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 greatest hesitation is the lack of comparison to other augmentation strategies. A reader cannot know if the proposed methodology outperforms contemporary augmentation techniques.
Other than that, the work seeks to solve an important clinical problem and proposes an innovative solution for improving synthetic augmentation for cases with rare pathologies/objects. The public release of real-world RFOs will also be of great interest to the community.
Review #2
- Please describe the contribution of the paper
The paper introduces SurgRFO, a two-stage synthetic augmentation pipeline that generates realistic RFO-positive surgical chest radiographs and shows that these synthesised positives improve downstream object detection performance, including evaluation on an external site.
- 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.Detecting critical RFOs is a safety-critical task where real positive data are rare; focusing on improving sensitivity/robustness is well aligned with clinical needs.
2.The proposed synthetic positives are used to improve multiple mainstream detectors, suggesting good practical utility and ease of adoption in existing clinical/engineering pipelines.
- 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 paper does not clearly state whether the data used to train the synthesis models (Stage 1/Stage 2) come strictly from the Site 1 training split, or whether Site 1 validation/test cases might be used during synthetic data generation. If the synthesis model is trained on test-patient images and then used to generate data for downstream evaluation, this would compromise the fairness of the internal assessment.
2.In the downstream detection experiments (Sec. 3.2), multiple metrics are reported (e.g., mAP, FROC, FNR, FOR, ACC), but their exact definitions and computation protocols in the detection setting are not clearly specified. Moreover, the overall experimental setup and evaluation procedure in this section are not sufficiently clear.
3.The data section states that Site 2 provides “An additional 20 critical RFO-positive cases,” yet the external results include metrics that depend on false positives (e.g., FROC). The authors should clarify whether the Site 2 external test set includes negative samples, and provide its size and class composition. If no negative cases are included, the paper should explain how these metrics are defined and computed for the external evaluation.
4.Table 4 reveals an interesting gap: in (a), clinicians give comparable quality/realism scores to real and synthetic images, yet in (b), they can still distinguish synthetic images with an accuracy of ~75%. Could the authors explain why the images receive similar quality/realism ratings despite being distinguishable?
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
None
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The authors’ work is meaningful and well-motivated; however, I have a few questions regarding the experimental section that would benefit from clarification.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The author did not fully or clearly address the weakness I highlighted earlier.
Review #3
- Please describe the contribution of the paper
The paper presents a two-stage compositional synthesis framework (SurgRFO) for generating intraoperative chest X-rays with critical retained foreign objects (RFOs) to address the scarcity of positive training data. Specifically, it decouples global background modeling using a diffusion-based foundation model from localized RFO synthesis via patch-based generation and conditional Poisson blending. In addition, the authors introduce the Intraoperative RFO Bench, a curated dataset of annotated surgical radiographs, and demonstrate that synthetic data augmentation improves downstream detection performance across multiple models.
- 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.Clinical relevance and problem importance The paper addresses the detection of critical retained foreign objects (RFOs), which are rare but high-risk events in intraoperative settings. This problem is highly relevant to patient safety and remains underexplored due to extreme data scarcity and subtle visual characteristics. The focus on such a clinically meaningful and challenging task significantly strengthens the impact of the work.
2.Well-motivated compositional synthesis framework The proposed two-stage framework separates global background generation from localized object synthesis, which aligns well with the inherent structure of intraoperative radiographs. This design enables better modeling of rare and small objects that are difficult to capture in fully end-to-end generative approaches. The use of conditional Poisson blending further ensures photometric consistency between synthesized objects and background, improving visual coherence .
3.Performance improvements across downstream task models The experimental results show that augmenting training data with synthesized samples leads to consistent improvements across multiple detection architectures, including Faster R-CNN, RetinaNet, and YOLOv8.
4.Inclusion of clinician-centered evaluation The incorporation of a radiologist-driven Turing-style evaluation adds a clinically meaningful validation of the generated images. The comparable quality ratings between synthetic and real images under blinded assessment support the plausibility of the synthesis results and enhance the translational relevance of the study .
- 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.Limited integration between stages The proposed method adopts a two-stage pipeline in which background generation and object synthesis are trained independently and combined through a post-hoc optimization process. However, this design prevents joint feature learning between the background (normal chest radiographs) and the synthesized objects (RFOs). As a result, the model is unable to capture globally consistent representations, which may limit both visual coherence and scalability of the framework.
2.Limited dataset size and insufficient generalization validation The dataset contains only 144 positive cases, and the external test set is very small (20 cases). While it is understandable that positive cases are inherently scarce, there exists a large amount of publicly available negative data. Incorporating external public datasets, particularly for downstream evaluation and Turing test analysis, would strengthen the conclusions and provide more convincing evidence of generalization.
3.Incomplete baseline comparisons The experimental comparisons are limited to RoentGen and DreamBooth, which do not fully represent the current state of the art. The absence of more recent and competitive baselines makes it difficult to assess the true advantage of the proposed method. Including stronger and more up-to-date generative models would provide a more comprehensive evaluation.
4.Evaluation metrics are not fully aligned with the task characteristics The use of FID and MS-SSIM for evaluating medical X-ray synthesis is debatable, as these metrics do not necessarily reflect clinical fidelity or the realism of localized lesions. The paper itself acknowledges that MS-SSIM is largely dominated by background similarity. More task-specific metrics would provide a more meaningful and clinically relevant evaluation.
5.Limited analysis of downstream task improvements The downstream evaluation aims to demonstrate that incorporating synthetic data improves detection performance. However, the analysis lacks depth. In addition to accuracy, aspects such as stability and robustness should be considered, for example by reporting standard deviations across multiple runs. Furthermore, evaluating performance under reduced base data (data ablation) could provide stronger evidence of the effectiveness of the synthesized data. Relevant related studies have demonstrated such evaluation strategies and could be referenced for comparison.
Layer Separation: Towards Adjustable Joint Space Width Images Synthesis (ACM MM 2025) https://doi.org/10.1145/3746027.3755407 Towards Generalizable Tumor Synthesis (CVPR 2024) https://openaccess.thecvf.com/content/CVPR2024/papers/Chen_Towards_Generalizable_Tumor_Synthesis_CVPR_2024_paper.pdf
6.Lack of discussion on X-ray imaging physics X-ray image synthesis differs fundamentally from natural image synthesis, as it is governed by specific imaging physics. The current approach, which relies on a “copy-paste”-style compositional strategy, may not fully respect the physical formation process of X-ray images (e.g., attenuation and projection). A more detailed discussion or justification of how the proposed method aligns with or approximates these physical principles would strengthen the work.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
This paper addresses a clinically important and challenging problem, namely the detection of rare critical retained foreign objects (RFOs) in intraoperative radiographs. The work is well-motivated, and the proposed framework provides a practical solution to the issue of extreme data scarcity by leveraging synthetic data generation. The introduction of a curated dataset and the consistent improvements observed across multiple detection models suggest that the approach has meaningful potential and practical value.
While the methodological novelty is somewhat limited, as the framework mainly combines existing techniques in a structured manner, the design is reasonable and tailored to the characteristics of the problem. The inclusion of clinician-based evaluation further strengthens the relevance of the work from a translational perspective.
There are, however, several aspects that could be further improved, including stronger baseline comparisons, more comprehensive validation, and more task-specific evaluation metrics. Addressing these points would make the contribution more convincing and robust.
Overall, despite some limitations in novelty and experimental depth, the paper presents a meaningful application with practical impact and a reasonably solid experimental demonstration. With further refinement, it has the potential to make a useful contribution to the field, and I consider it suitable for borderline acceptance.
- 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 paper is overall well-structured and addresses a meaningful research problem with clinical relevance. The rebuttal addressed several important concerns raised during review. Therefore, I believe the revised paper is marginally above the acceptance threshold for MICCAI.
Author Feedback
We thank the reviewers and meta-reviewer for their constructive comments. All reviewers (R1–R3) acknowledged the clinical importance of rare, safety-critical RFO detection. R2, R3 recognized SurgRFO as a meaningful two-stage synthetic augmentation framework, noting its downstream detection gains (R2, R3), robust synthetic image evaluation including clinician assessment (R1, R3), external validation (R1, R2), code release, and planned release of intraoperative RFO dataset (R1, R3). Below, we address the reviewers’ concerns. R1, Meta-R (Relation to prior RFO synthesis work and novelty): We agree that Zhou et al. (MICCAI DEMI 2026) is relevant and should be contextualized. That work releases CT-based physically synthesized RFO images, rather than real intraoperative radiographs with critical RFOs. Its pipeline inserts RFO meshes into relatively clean CT volumes from normal patients and projects them to X-rays, which is valuable but differs from our target setting: cluttered surgical chest radiographs acquired during real intraoperative workflows. SurgRFO is designed around this domain gap. We curate and release, to our knowledge, the first publicly available real de-identified intraoperative surgical chest radiograph dataset dedicated to critical RFO detection with bounding-box annotations, distinguishing it from prior physically synthesized CT-projection datasets. In the revision, we will add this work to the Introduction, clarify this distinction, and include a contextual comparison. R1, Meta-R (Dataset release): We confirm that, upon acceptance, our public release will include de-identified intraoperative RFO radiographs, annotations, and all synthetic SurgRFO images used in the paper. The earlier GitHub statement was imprecise and has been corrected. We also provided example real and synthetic images in the earlier repository to demonstrate the data format and visual quality. R1, R2, Meta-R (Cohort details, preprocessing, and leakage prevention): We will add a dedicated paragraph detailing cohort composition, labeling, and experimental protocols. The Site-1 cohort contains 144 critical RFO-positive cases and 944 RFO-negative cases, with one case per patient. Positive and negative cohorts were split at the patient level using the same 6:1:3 train/validation/test protocol, ensuring no patient overlap. We will report exact positive/negative counts for each split. All synthesis components, including background adaptation, RFO appearance modeling, placement-prior estimation, and blending-parameter selection, used only the training split; validation/test images were never used for synthesis training or tuning. Detectors were trained on the training split, selected using validation data, and evaluated only on held-out internal and external cohorts. All images followed the same preprocessing pipeline. We will also clarify labeling criteria/workflow, detector training details, and hyperparameters; implementation details are available in the earlier released code. R1, R3, Meta-R (Compare with simpler augmentations): We agree that stronger augmentation comparisons are important. Both real-only and SurgRFO-augmented detector training used the same standard data augmentations, including random flipping and intensity normalization. Thus, the reported gains are not due to basic augmentation being applied only to SurgRFO. To further isolate the contribution of the proposed synthesis strategy, we will add a simple RFO copy-paste/injection baseline using the same number of synthetic-positive samples, allowing direct comparison between SurgRFO and increased synthetic sample count alone. We will clarify the full augmentation protocol. R1, R3, Meta-R (Statistical analysis): We agree that uncertainty analysis is important in this rare-event setting. We will add patient-level bootstrap 95% confidence intervals for the primary downstream metrics using fixed held-out test predictions. These intervals will be reported in the main tables.
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.
Paper summary: This paper introduces SurgRFO, a two-stage synthetic augmentation pipeline that generates realistic RFO-positive surgical chest radiographs and shows that these synthesised positives improve downstream object detection performance, including evaluation on an external site.
Strengths: This work addresses a clinically important and safety-critical problem. A key strength is the introduction and public release of a rare clinical radiograph dataset containing critical RFO cases, which represents a valuable contribution to the field. The proposed well-motivated compositional synthesis framework, which separates radiograph generation from localized RFO synthesis, is technically sound and improves the fidelity and realism of synthesized objects. This is further supported by robust evaluation of synthetic image quality through blinded clinician assessment, quantitative distribution analysis, and external validation. Importantly, the synthetic data consistently improve downstream detection performance across multiple mainstream architectures (e.g., Faster R-CNN, RetinaNet, YOLOv8), demonstrating strong practical utility, ease of adoption in existing clinical pipelines, and alignment with the clinical need to improve sensitivity and robustness in rare-event detection.
Weaknesses:
-
First, the novelty is somewhat weakened by substantial overlap with prior work on synthetic augmentation for critical RFO detection, yet the manuscript does not adequately contextualize or compare against these closely related methods.
-
In addition, the downstream detection gains are not compared against alternative synthetic generation or simpler augmentation strategies, leaving it unclear whether the improvements stem from the proposed method itself or simply from increased training data.
-
The paper also lacks important dataset and experimental details, including cohort composition, labeling protocols, train/test separation for synthesis, preprocessing consistency, and external cohort setup - which raises concerns about reproducibility and evaluation fairness.
-
Furthermore, downstream performance is reported without statistical significance analysis or measures of uncertainty, despite relatively small positive sample sizes, limiting confidence in the robustness of the reported gains.
Decision Rationale: The paper has clear strengths, but several substantive concerns remain unresolved, particularly regarding the limitations above. These points appear addressable through author clarification. I therefore recommend Invite for Rebuttal.
-
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Thank you for the constructive rebuttal for addressing the reviewers’ concerns. The discussion improved confidence in both the technical contribution and the clinical relevance of the work.
The Reviewers acknowledged that the paper tackles an important clinical problem and proposes an innovative approach for synthetic augmentation in the setting of rare pathologies and objects. This is an interesting contribution for the MICCAI community.
One remaining concern is the limited comparison against other contemporary augmentation strategies, which makes it difficult to fully position the proposed method within the broader augmentation literature. While this limitation should be acknowledged more clearly, the current evidence is sufficient to support the paper’s main claims. The planned public release of real-world RFO data is also expected to provide meaningful value to the community.
For the camera-ready version, I recommend authors to strengthen discussion of related augmentation approaches and clarify the scope of claims accordingly.
I support the acceptance of this paper.
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.
This manuscript presents a method to generate synthetic XRay images with foreign objects to aid in the development of a method to recognise foreign objects XRay images. This addresses a clinically important dataset and the authors will also release this dataset.
The reviewers found the method to be technically sound and interesting. The limitations were limited novelty, and no comparison with other augmentation strategies.
Although there is some overlap in methodology with prior works, this manuscript is evaluated on real images with foreign objects whereas other methods generated synthetic evaluation image.
I am accepting this manuscript because on balance I find the novelty an important step change the other weaknesses (lack of comparison with other data augmentation methods) are counter balanced by the strengths of the work.
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 reviewers’ concerns have been successfully addressed in the rebuttal, making the work a solid contribution especially for what concerns methodological applications in MIC.
