List of Papers Browse by Subject Areas Author List
Abstract
Background: Superficial radiotherapy still relies on CT simulation for dose calculation, adding resource burden and imaging radiation for benign or shallow disease. Optical surface scanning is non-ionizing, but surface-to-CT synthesis is ill-posed because one surface can match multiple internal anatomies.
Methods: We propose SuRe-Flow (Surface Retrieval Flow), a retrieval-augmented rectified-flow framework for superficial synthetic CT generation. A contrastive surface encoder retrieves topologically matched bulk anatomical priors to reduce 1-to-N ambiguity; ControlNet injects these priors into a pretrained RF backbone, and a dose-aware frequency loss preserves tissue-interface detail for electron dosimetry.
Results: Trained on 20,022 patients and evaluated on 1,000 patients, SuRe-Flow outperformed adversarial and rectified-flow baselines, achieving 5.0 HU MAE in the superficial therapeutic zone (0-20 mm). Treatment-planning-system dosimetry yielded a 96.7% 3D Gamma Index passing rate (2% 2mm).
Conclusion: SuRe-Flow enables dosimetrically accurate surface-conditioned CT synthesis, supporting the feasibility of CT-free superficial radiotherapy.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3125_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiuCon_SuReFlow_MICCAI2026,
author = { Liu, Cong AND Ni, Xinye AND Xie, Kai},
title = { { SuRe-Flow: Surface-to-CT Synthesis via Retrieval-Augmented Rectified Flow for CT-Free Superficial Radiotherapy } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16890},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes SuRe-Flow, a retrieval-augmented rectified flow framework that enables CT-free superficial radiotherapy by synthesizing volumetric CT from optical surface scans. The core contributions are: (1) a contrastive representation learning module that retrieves topologically matched anatomical priors from a large database to resolve the ill-posed 1-to-N mapping ambiguity inherent in surface-to-CT synthesis; (2) a retrieval-conditioned flow mechanism that injects these priors via ControlNet to disentangle conflicting ODE trajectories in rectified flow; and (3) a dose-aware frequency penalty that explicitly regularizes high-frequency boundaries to preserve tissue interfaces critical for electron beam dosimetry. The work demonstrates clinical feasibility through large-scale validation (20,022 training/1,000 test patients) and clinical TPS evaluation, achieving 5.0 HU MAE in the 0–20mm therapeutic zone and 96.7% Gamma Index passing rate.
- 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.
Novel solution to ill-posed inverse problem: The retrieval mechanism elegantly addresses the fundamental 1-to-N ambiguity (single surface → multiple valid internal anatomies) that causes mode averaging in standard conditional flow models. The contrastive learning approach (DSC 0.9442 for Top-1 retrieval) effectively anchors generation to anatomically plausible priors. Strong clinical motivation and validation: Targets a genuine unmet need—eliminating unnecessary CT scans and ionizing radiation exposure for superficial radiotherapy patients (keloids, skin cancer). The dosimetric evaluation in Monaco TPS using 3D Gamma Index (2%/2mm) provides clinically relevant validation beyond standard image quality metrics. Large-scale robust evaluation: Training on 20,022 patients from 100+ centers across 16 countries ensures broad anatomical diversity and generalizability. The depth-stratified analysis (0–20mm, 20–50mm, 50–100mm) demonstrates specific efficacy in the critical therapeutic zone. Technical integration: Sophisticated combination of rectified flow (stable generative prior), ControlNet conditioning, and focal frequency loss specifically designed to preserve sharp density gradients required for accurate electron density heterogeneity corrections. Quantitative superiority: Substantial performance gains over baselines (MAE 5.0 HU vs. 30.2 HU for standard RF in Zone 1; 96.7% vs. 89.4% Gamma passing rate), establishing genuine clinical feasibility.
- 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.
Limited generalizability beyond superficial targets: The method is specifically optimized for 0–20mm depth (electron beam range). The retrieval mechanism and frequency penalty may not translate to deeper anatomical targets or other imaging modalities without fundamental redesign. Database coverage concerns: While the 20,022-patient database is large, performance depends on retrieving sufficiently similar anatomies. Patients with unusual body habitus, pathological deformities, or surgical alterations may retrieve suboptimal priors, though the paper shows robustness via Top-5 neighborhood sampling during training. Missing comparisons: While compared to MAISI-v2 and a 3D GAN, the paper lacks comparison against other recent medical image synthesis approaches (e. g. , score-based diffusion models, VQ-GANs, or other retrieval-augmented generation methods outside medical imaging). Computational overhead: The paper does not discuss inference latency or computational cost associated with the retrieval database queries (FAISS) and ControlNet architecture, which is relevant for real-time clinical workflow integration with optical surface scanners. Clinical validation depth: While dosimetric accuracy is shown, the study does not assess inter-observer contouring variability on synthetic vs. real CTs, or downstream clinical outcomes (tumor control, toxicity), though these are beyond the scope of a methodological 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 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.
(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 paper presents a technically sophisticated and clinically significant solution to surface-to-CT synthesis. The retrieval-augmented approach directly addresses the ill-posed nature of the inverse problem through a well-designed contrastive learning mechanism, while the dose-aware frequency penalty demonstrates deep understanding of radiotherapy physics requirements. The large-scale validation (20k+ patients) and clinical TPS evaluation provide compelling evidence of efficacy. The 96.7% Gamma pass rate establishes genuine clinical feasibility for eliminating CT scans in superficial radiotherapy, representing both technical novelty and potential for real-world patient benefit (reducing unnecessary radiation exposure).
- 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 #2
- Please describe the contribution of the paper
The paper proposes SuRe-Flow, a framework for generating synthetic CT volumes directly from body surface information for superficial radiotherapy planning. The method combines a retrieval module that finds anatomically similar prior cases with a rectified flow backbone, and further adds a dose-aware frequency loss to preserve shallow tissue boundaries that matter for electron dose calculation. The overall goal is to support a CT-free workflow for superficial treatments while maintaining acceptable dosimetric accuracy.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper has several strengths. The most obvious one is the clinical motivation. For superficial radiotherapy cases, avoiding CT simulation could reduce unnecessary imaging burden and simplify workflow, so the application itself is meaningful.
Another strength is that the method is well aligned with the problem. Surface-to-CT synthesis is highly ambiguous, and the retrieval module is a reasonable way to provide anatomical priors instead of relying only on a generic generative model. The combination with a rectified flow backbone is also interesting and technically relevant.
The experimental setup is also solid. The study uses a large dataset and evaluates not only image metrics but also dosimetric Gamma Index results in a treatment planning system. This makes the paper stronger than many synthesis works that only report visual similarity metrics.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The main concern is the realism of the input setting. The paper aims at generating CT from optical surface scans, but the input surface appears to be derived from CT body contours rather than real scanner data. In practice, optical scans may contain noise, missing regions, motion, clothing artifacts, or posture mismatch. Because of this, it remains unclear how well the method would transfer to a real clinical workflow.
Another concern is the retrieval component. Since the method relies on retrieving similar prior cases from the same large dataset, it would be helpful to clarify how very close matches were handled. A comparison with a simple retrieval-only baseline would also help show how much of the improvement comes from the generative model beyond the retrieved prior itself.
The clinical validation is promising but still somewhat limited. The paper reports Gamma Index results, which is a strong point, but it would help to know more about the treatment setups that were evaluated and whether results are consistent across different body locations. It would also be useful to see how the method performs in more difficult or less common cases.
Finally, although the overall design is reasonable, the paper does not discuss practical efficiency such as inference time, hardware requirements, or workflow integration. These aspects would be useful for judging real clinical usability.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
Thank you for the interesting submission. It would be helpful to clarify whether the authors plan to release code or trained models upon acceptance, as this would further improve reproducibility and community impact.
- 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 meaningful clinical problem and proposes a method that is both relevant and reasonably well motivated. Reducing or removing the need for CT simulation in superficial radiotherapy would be valuable in practice, which gives the paper clear translational relevance.
The technical idea is also solid. Using retrieval to provide anatomical priors for an otherwise ambiguous surface-to-CT generation task is a sensible design choice, and combining this with a rectified flow backbone is interesting. The experimental study is another positive point, especially the use of a large multi-center dataset and the inclusion of dosimetric Gamma Index evaluation rather than relying only on image similarity metrics.
At the same time, I do have some reservations, which is why I did not score it higher. The main issue is that the input surface seems to be derived from CT contours rather than real optical scanner data, so the gap to an actual clinical workflow remains unclear. I also think the retrieval setup and practical deployment aspects could be discussed in more detail.
Overall, I found the paper to be a useful and timely contribution with clear potential impact, even though some limitations still need to be clarified.
- 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
Retrieval-augmented rectified flow framework for surface-to-CT synthesis using contrastive anatomical prior injection Entropy-gated ControlNet conditioning that narrows the 1-to-N ODE trajectory solution space via retrieved CT latents Dose-aware focal frequency penalty targeting high-frequency tissue interface fidelity Depth-stratified evaluation demonstrating dosimetric viability via Monaco TPS Gamma Index
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The clinical motivation is well-grounded in radiotherapy physics, and the focus on the superficial zone (0–20 mm) makes the problem both tractable and clinically meaningful. The inclusion of end-to-end dosimetric validation in a treatment planning system (3D Gamma 2%/2mm) is a significant strength and demonstrates real clinical relevance beyond image-level metrics. The scale of training and evaluation (20k+ patients across multiple centers and countries) is substantial and supports claims of robustness and generalizability. The ablation study clearly demonstrates the contribution of each module
- 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 retrieval module is only validated using external body contour similarity and does not demonstrate alignment of internal tissue density, leaving its core purpose insufficiently supported. 2.The impact of retrieval is not rigorously validated — no ablation with random or mismatched priors is provided, making it unclear whether retrieval is actually responsible for the observed gains. 3.The contrastive learning setup is weak, as positive pairs are generated via augmentations of the same sample rather than anatomically similar patients, which does not guarantee meaningful retrieval behavior. 4.Baseline comparison is limited for a paper of this scale, with no inclusion of recent diffusion-based or anatomy-conditioned synthesis methods, weakening the strength of the empirical claims.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The work is strongest as a clinically motivated study with dosimetric validation rather than a purely methodological contribution; positioning it more explicitly as a clinical translation paper would improve clarity and differentiation. The assumption that CT-derived body contours approximate real optical surface scans is not validated; differences in noise, occlusion, and resolution could significantly affect performance, and no robustness analysis is provided. A comparison with atlas-based sCT methods, which are standard in radiotherapy workflows, is missing and would provide a stronger clinical baseline.
- 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?
Clinically meaningful with strong scale and dosimetric validation, but key components (especially retrieval) are not rigorously validated and some assumptions remain unverified. Overall promising but not fully complete, hence borderline accept.
- 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
Author Feedback
To clarify, we represent the surfaces as voxelized occupancy masks. This allows optical scanner point clouds and CT-derived external contours to share the same downstream model input after voxelization.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper introduces a contrastive surface-based retrieval module that identifies and injects topologically matched anatomical priors during generation. The overall writing is good, and the method is novel. Beyond the reviewers’ comments, the authors should further discuss how the method ensures geometric fidelity for individuals. Although the retrieval method is used to improve it, this may not be able to depict the lesion with large inter-variations. It would also be beneficial to provide more details on clinical tolerance for the realism of the generated CT images.