List of Papers Browse by Subject Areas Author List
Abstract
Glioblastoma is an aggressive brain tumor requiring precise treatment planning. Magnetic resonance imaging (MRI) is essential for pre-operative assessment, surgical resection planning, and post-operative monitoring. Therefore, generating post-operative MRI from pre-operative MRI can assist neurosurgeons in many ways, such as predicting surgical outcomes and guiding treatment planning. However, generating post-operative MRI from pre-operative MRI is challenging, as the resection extent depends on tumor location and infiltration to minimize potential complications, necessitating consideration of surgical outcomes based on tumor location and shape. Furthermore, post-operative MRI differs significantly from pre-operative MRI due to structural and visual changes, such as tissue shift, edema, hemorrhage, and the resection region. To address these challenges, we propose a novel post-operative MRI generation method that generates post-operative MRI from pre-operative MRI using tumor-aware visual in-context learning. Specifically, we provide explicit visual instruction for generating post-operative MRI from pre-operative MRI, improving the capture of structural changes. To consider tumor-specific post-operative outcomes, we propose tumor-guided retrieval, which retrieves the tumor case most similar to the query pre-operative MRI, and a tumor-aware prompt adapter that integrates tumor resection and anatomical structure information. Our proposed method achieves superior performance on publicly available dataset and is the first to generate post-operative MRI from pre-operative MRI, introducing a new approach to improving patient prognosis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3978_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{KanBog_PretoPost_MICCAI2025,
author = { Kang, Bogyeong and Park, Sang-Jun and Lim, Minjoo and Kang, Myeongkyun and Heo, Keun-Soo and Oh, Ji-Hye and Lee, Hyun Jung and Kam, Tae-Eui},
title = { { Pre-to-Post Operative MRI Generation with Retrieval-based Visual In-Context Learning } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15960},
month = {September},
page = {537 -- 547}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors predict the post-operative MRI based on the pre-operative MRI of brain tumors. The authors claim that their approach outperforms several baseline methods in this task.
- 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.
- Figure 1 is very well designed.
- The authors provide a solid selection of baselines.
- The pure results look promising (assuming the missing statistics work 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.
-
For me, the problem that the authors aim to solve is not apparent. The post-operative image is acquired in any case, and it is unclear how estimating it could help plan surgical outcomes, since these depend on the surgery itself. Even the hand-picked example in Figure 2 (second row) shows large differences between the ground truth and the proposed method.
-
In Tables 1 and 2, the authors report results with four decimal places but do not provide any standard errors or statistical significance tests to assess the variability within each method.
-
No direct image similarity metrics such as mean squared error are reported. It is unclear why this was omitted, especially since the goal is not to generate realistic images in general, but to predict the patient-specific post-operative image. At the very least, a comparison of the resulting resection cavities, for example using Dice scores, would be necessary.
-
- 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 provide sufficient information for reproducibility.
- 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/2025/en/REVIEWER-GUIDELINES.html
Besides the listed critiques, I want to encourage the authors to continue in this research direction. I believe that reframing the approach in a more probabilistic manner, for example by predicting the probability of resection for each voxel, could generate significant clinical value, which I think is missing in the current form.
- 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?
In my opinion, the purpose of the application is unclear. It is not evident how an estimated post-operative image would support clinical decision-making. Therefore, it is not surprising that this is the first work addressing the task in this way.
Besides that, the inclusion of additional datasets would be necessary to establish the general validity of the proposed method.
Additionally, the results do not follow fundamental statistical conventions, which makes them hard to interpret. Important metrics are missing, and the reproducibility of the study is questionable, as no code has been provided.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
I still disagree with the purpose the authors claim. Unless the model is trained to account for variations in therapy or treatment, any “simulated” post-operative MRI will merely reflect noise rather than meaningful surgical outcomes. Without explicitly incorporating those treatment options into the training process, the generated images are effectively random and can’t provide reliable guidance for clinicians or patients.
The absence of statistical metrics, such as error bars or p-values, prevents any meaningful validation of the findings. Ignoring my request for these analyses undermines confidence in the results’ significance.
Omitting key quantitative metrics like MSE and failing to release the code raises serious concerns about transparency and reproducibility.
Review #2
- Please describe the contribution of the paper
This paper proposes a diffusion-based model for post-operative MRI synthesis, incorporating context-guided generation through paired pre- and post-operative data, along with tumor-guided retrieval to enhance structural consistency.
- 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 introduces an interesting approach that retrieves similar MRI pairs to serve as guidance during the diffusion process, enhancing the generation of post-operative images. The tumor segmentation results are effectively used as a basis for retrieval, and a tumor feature prompt is constructed by combining information from both synthetic (make-up) images and real segmentation results. The evaluation and comparisons with existing methods are thorough and clearly presented, contributing to the paper’s overall clarity and reproducibility.
- 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 paper does not provide sufficient details regarding the bounding box generation or placement, which is important for understanding how tumor regions are localized and utilized in the model. Clarifying this would improve transparency and reproducibility.
In the ablation study, it would strengthen the paper to compare different backbone architectures for the tumor-aware prompt adapter. Such comparisons could provide more insight into its effectiveness and robustness.
It is unclear whether the Top-K retrieval is treated as distinct training pairs or part of the same training instance. Additionally, the choice of K = 15 appears arbitrary—justifying this selection or including an ablation on different K values would enhance the credibility and completeness of the experimental analysis.
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
The font size and style in the figures should be made consistent throughout the paper. This would improve the visual quality and overall presentation.
The paper could be further strengthened by exploring alternative retrieval strategies, such as Dice similarity, Euclidean distance (ED), or Hausdorff distance-based retrieval. Evaluating different similarity metrics could provide deeper insight into the robustness and effectiveness of the retrieval-guided diffusion framework.
- 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 — could be accepted, 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 novel idea by introducing additional information to regularize the diffusion process. The use of IoU-based retrieval is both clever and efficient, allowing the model to incorporate structurally relevant guidance in a lightweight manner.
- Reviewer confidence
Very confident (4)
- [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
In this paper, the authors propose a method to generate MR images of patients who underwent glioblastoma surgical resection from a pre intervention MRI and a corresponding segmentation mask. The method relies on a database of paired pre-operative MR images, post surgical resection MR images, and corresponding post-op segmentation mask. The author introduce a tumor-aware prompt adapter and a tumor-guided retrieval technique to generate the post-op MR from the pre-op MR. The tumor-aware prompt adapter focuses on considering post operative outcomes specific to tumor location and shape. The tumor-guided retrieval technique enhances the capture of tumor-specific post-operative structural changes. The method is build and tested on a public dataset (LUMIERE) and compared against 5 other diffusion based image editing methods. An ablation studies highlights the importance of the tumor-aware prompt adapter and of the tumor-guided retrieval features.
- 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 is well written and easy to follow. The method proposed is novel and clearly explained. The ablation study showing the added value of the tumor-aware prompt adapter and of the tumor-guided retrieval features supports the paper claims.
- 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.
My major concern is the weak evaluation of the method. (1) the evaluation is limited to the same dataset used for training. It would have been more convincing to extend it to an external dataset (maybe BRATS?). (2) a visual qualitative evaluation of the results by a trained neurosurgeons to assess if the images are realistic would have also been interesting. Another concern is that the method is 2D without the author explaining this choice. As I have doubt about the use in clinical setting of such a 2D method, I wonder how would the translation to 3D work.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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/2025/en/REVIEWER-GUIDELINES.html
Minor comment:
Training/testing split: since 2D slices are extracted, I think that it is important to highlighted that the split does maintain the distinction between different patients as I hope it does!
It is not clear how the tumor-guided retrieval technique works. Are the images in the database all registered somehow? How is the queried image registered to the database?
How does the axial plane alignment works? it is not clear.
Not enough details are given on the segmentation masks needed as input: It seems to be multi-labels from the figure but it is not mentioned in the text. Which labels are present? are they needed, are they treated differently?
For the tumor-guided retrieval: IoU is used to retrieve a similar MRI triplet. But what are the range of IoU observed? What is a good IoU? The authors seems to claim that this is a key component of the method, but how sensitive is the method to the IoU values? How is the database used for retrieval built? Is it done randomly or you ensure that you had a complete representation of all possible tumor locations in the database?
How are the top-k = 15 retrieved images used during training?
p5: Metrics like SSIM, MS-SSIM, PSNR, and LPIPS are used without introduction of the terms. Same for the competing methods: DDPM, PBE, VISII
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Good and well written paper presenting a novel and clearly explained approach, with an ablation study supporting the paper claims, but with a limited evaluation.
- 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.
Authors answered most of the important questions raised by the reviewers, and most of the unclear points are now clarified.
Author Feedback
We thank Reviewers 1, 2, and 3 for their helpful feedback.
Purpose and impact of our study (R3): Our goal is not to replace post-operative imaging, but to support clinicians and patients in developing a comprehensive treatment plan for glioblastoma (GBM) by simulating post-operative MRI before any treatment begins. Since GBM treatment includes not only surgery but also chemotherapy and radiotherapy, it is important to design an appropriate treatment plan that considers whether and when to administer each of these treatment options.
While neurosurgeons may currently plan treatment based on pre-operative MRI alone, our post-operative MRI simulation can provide valuable additional guidance during treatment planning. It can offer visual access to expected surgical outcomes, helping to assess potential surgical impact and support decision-making about whether to proceed with surgery. It can also help patients better understand the proposed treatment strategy by providing the expected surgical outcomes.
Furthermore, while the current study focuses on post-operative MRI generation, it is a necessary foundational step toward more patient-specific treatment planning and prognosis prediction for GBM. We plan to extend this study by generating longitudinal follow-up MRI scans based on post-operative MRI to support such treatment planning.
Evaluation metric and Dataset (R2&R3): We used four metrics: SSIM and MS-SSIM for structural similarity, and PSNR and LPIPS for perceptual similarity. Since PSNR is derived from MSE and offers a more interpretable measure on a logarithmic (dB) scale, we reported PSNR instead of MSE. We plan to include MSE, Dice score, additional datasets, and expert assessments in future work. The training and test sets consist of entirely different patients.
Training strategy and selection of k in retrieval (R1&R2): During training, we retrieve the top-15 most similar triplets for each query. From these retrieved triplets, one triplet is randomly selected in each batch and used as the instruction. This strategy ensures diversity and prevents overfitting caused by repeatedly using the same instruction during training. Our preliminary experiments showed that k=15 performed best among 5, 10, 15, and 20, with smaller k causing overfitting due to limited diversity and larger k introducing too much variation in tumor size and location.
Details on bounding box (R1): Given a query pre-operative MRI and its segmentation mask, we compute an enclosing bounding box and create a bounding box mask with 1 inside the box and 0 outside. To obtain Q_comp, the region corresponding to 1 in the bounding box mask is taken from the retrieved instruction’s post-operative MRI, while the region corresponding to 0 is preserved from the query pre-operative MRI. This strategy enables the generated post-operative MRI to better reflect tumor-specific resection outcomes, as supported by Table 2.
Comparison with other prompt adapters (R1): PBE, InstructPix2Pix, VISII, and ImageBrush in Table 1 use different prompt adapters. Thus, Table 1 provides an indirect comparison, where our tumor-aware adapter performs best, showing its effectiveness.
Clarifications of Tumor-guided Retrieval (R2): For each query pre-operative MRI, we compute IoU between its mask and those of all triplets in the database, with observed IoU values ranging from 0.0 to 0.93. To reflect tumor location along the axial plane, we restrict candidates to slices with similar indices and select the top-k with highest IoU. We found that multi-label IoU greatly increased computation time without performance improvement, so we use binary masks for retrieval.
Future work (R1&R2&R3): We adopt a 2D slice-based approach to leverage pre-trained 2D foundation models for efficient training. While our current approach can reconstruct 3D volumes effectively by stacking 2D slices, we plan to explore full 3D generation, alternative retrieval strategies, and statistical analyses in future work.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
The paper presents an interesting application of post-operative 2D brain MRI synthesis. To strengthen the work, the authors should include evaluation on an external dataset and consider extending the method to 3D.