List of Papers Browse by Subject Areas Author List
Abstract
Glioblastoma is a highly malignant brain tumor marked by rapid progression and high mortality, making patient-specific treatment strategies essential for improving prognosis. Longitudinal post-operative magnetic resonance imaging (MRI) plays a key role in monitoring surgical outcomes and planning subsequent therapy. Despite significant advances in generative models, longitudinal post-operative MRI generation remains challenging, as it requires reflecting clinical preferences that align with patient-specific post-operative outcomes. Such clinically preferred images must accurately reflect long-term post-operative trajectories that depend on the extent of resection (determined by tumor location and infiltration) and progressively evolve across follow-up time points. However, existing generative models often fail to capture these clinical preferences, leading to misalignment between generated images and clinically meaningful outcomes. To address these challenges, we propose TADPO, a novel Tumor-Aware Direct Preference Optimization (DPO) framework that aligns longitudinal post-operative MRI generation with clinical preferences using pre-operative MRI. Specifically, TADPO leverages DPO to generate post-operative MRIs aligned with preferred outcomes across multiple follow-up time points. We also introduce a tumor-aware optimization strategy that prioritizes tumor regions and their boundaries, improving anatomical fidelity of post-operative tumor representations and enhancing alignment with clinically meaningful preferences. TADPO demonstrates superior alignment with clinical preferences compared to state-of-the-art generative methods, highlighting its potential to support personalized treatment planning.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2027_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_TumorAware_MICCAI2026,
author = { Kang, Bogyeong AND Lim, Minjoo AND Nam, Hyeonyeong AND Heo, Keun-Soo AND Liu, Mingxia AND Kam, Tae-Eui},
title = { { Tumor-Aware Direct Preference Optimization for Longitudinal Post-operative MRI Generation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16887},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes TADPO, a novel framework for longitudinal post-operative MRI generation that incorporates clinical preference alignment via DPO. First，the authors construct a preference dataset without human annotation using ref model. Then they introduce a tumor-aware weighting strategy based on SDF to emphasize tumor boundaries and enforce fidelity in tumor core regions. Finally, they perform longitudinal generation conditioned on pre-operative MRI with temporal prompts. Overall, this paper presents an interesting clinical problem and provides an appropriate solution.
- 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 major strengths of this paper are summarized as follows: 1.Preference-aligned longitudinal MRI generation is a interesting problem formulation. It is also a conceptual advancement beyond standard conditional generation. 2.A good attempt to adapt DPO to longitudinal medical image generation with a novel preference construction strategy. 3.Tumor-aware weighting strategy is well-motivated and clinically meaningful.
- 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 major weaknesses of this paper are summarized as follows: 1.The core assumption of the proposed framework is that ground-truth images correspond to “preferred” samples, while generated images are treated as “dispreferred. ”The assumption is not sufficiently justified in paper. It remains unclear whether the model truly learns preference alignment beyond enforcing consistency with ground truth, and further evidence or analysis is needed to support this claim. 2.Tumor evolution is a 3D spatiotemporal process, while the proposed method operates on 2D slices. 3.The dataset is split at the slice level for training the reference model and DPO model. This may introduce information leakage, as slices from the same patient could appear in both training subsets. A patient-level split would be more appropriate to ensure a fair evaluation. 4.The proposed loss function introduces some hyperparameters. There needs a sensitivity analysis or discussion. 5.Equations (2) and (4) appear to contain inconsistencies or possible errors in their formulation.
- 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?
While the presented work is well motivated and explores an interesting research direction, I still have several concerns regarding the methodological design and experimental validation. In particular, the use of 2D slice-based generation limits its clinical applicability, as tumor progression is inherently a 3D process. Ultimately, these issues lead me to give a weak accept recommendation.
- 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
This paper proposes Tumor-Aware Direct Preference Optimization (TADPO) for generating longitudinal post-operative MRI from pre-operative MRI. The main contributions are: (i) an automatic preference-pair construction strategy where ground-truth follow-up MRIs are treated as winning samples and reference-model-generated MRIs are treated as losing samples, (ii) a tumor-aware reweighting of the DPO objective using an SDF-based weighting map that emphasizes tumor boundaries and tumor interiors, and (iii) a conditional diffusion framework that incorporates both follow-up week information and pre-operative segmentation. Empirically, the method reports improved performance over several diffusion/image-editing baselines on the LUMIERE dataset, especially under tumor-region evaluation. That said, the contribution is better characterized as an application-oriented combination of DPO and tumor-aware spatial reweighting for longitudinal medical MRI generation, rather than a fundamentally new preference-learning formulation.
- 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 paper addresses a clinically meaningful and technically challenging task: longitudinal post-operative MRI generation, where both temporal progression and post-surgical tumor dynamics matter. 2.Introducing preference alignment into medical image generation is interesting, especially under the framing of clinical preference alignment for future follow-up MRI generation. 3.The tumor-aware reweighting is the most convincing technical idea in the paper. Emphasizing tumor boundaries and tumor interiors is better motivated than applying preference optimization uniformly over the whole image. 4.The evaluation is at least partially aligned with the paper’s claims. In addition to whole-image metrics, the paper reports tumor-region metrics, which are more relevant to the proposed tumor-aware alignment objective. 5.The ablation studies are relatively coherent. The paper examines the effect of DPO, vanilla vs. tumor-aware DPO, spatial and temporal prompts, and the data split ratio between reference-model pretraining and DPO fine-tuning.
- 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 main issue is that the notion of “preference” is not truly established as clinical preference in the strict sense. In practice, the paper uses ground-truth follow-up MRIs as winning samples and reference-model-generated MRIs as losing samples, which is closer to GT-vs-generated supervision than to a genuine human/expert preference dataset. As a result, it is unclear why DPO is necessary here, or how it is fundamentally different from a more direct supervised or weighted diffusion objective. 2.The tumor-aware weighting is intuitive but still fairly heuristic. The paper does not fully justify why an SDF derived from the pre-operative mask is sufficient to define the most important regions for future post-operative progression, and it does not compare against a harder tumor-only optimization variant that would better justify the proposed soft spatial weighting. 3.There is no comparison against other preference optimization or RL-based alignment methods. Therefore, it is difficult to determine whether the gain comes from DPO itself or simply from tumor-focused reweighted fine-tuning. 4.The baseline selection is not fully convincing. Most baselines are general diffusion or image-editing methods, whereas the task is highly specific to longitudinal post-operative MRI generation and clinical progression modeling. More task-relevant baselines would strengthen the empirical comparison. In addition, stronger lesion-aware/post-training medical diffusion baselines should be considered; for example, recent lesion-aware post-training work has demonstrated lesion-focused post-training on a substantially larger dataset. (Lee, Junhyeok, et al. “Lesion-Aware Post-training of Latent Diffusion Models for Synthesizing Diffusion MRI from CT Perfusion.” International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2025.)
- 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.
(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?
My recommendation is driven by the gap between the paper’s interesting problem setting and the strength of its actual evidence. The idea of introducing preference alignment into longitudinal post-operative MRI generation is interesting, and the tumor-aware reweighting of the DPO objective is well matched to the task. However, the claimed notion of “clinical preference alignment” is in practice based on automatically constructed pairs of ground-truth MRIs and reference-model-generated MRIs, which makes it much closer to GT-guided alignment than to genuine clinical preference learning. As a result, the paper does not convincingly establish why DPO is necessary here, or what fundamental advantage it has over a more direct region-weighted supervised diffusion objective. In addition, the work does not compare against alternative preference-optimization or RL-based alignment methods, and the baseline set is only partially task-appropriate.
- 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
The paper proposes TADPO, a tumor-aware Direct Preference Optimization framework for generating longitudinal post-operative MRIs from pre-operative scans. The key challenge is incorporating clinically meaningful factors, such as extent of resection and tumor progression over time, which are not well captured by standard generative models. To address this, the authors use a publicly available longitudinal dataset and apply DPO framework to align outputs with clinically realistic follow-up images, while introducing a tumor-focused optimization strategy that emphasizes relevant 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.
The proposed method effectively incorporates clinically meaningful factors like tumor resection and progression, producing outputs that are noticeably more clinically plausible than existing generative models, as demonstrated by the experiments. The tumor-aware DPO framework is a well-motivated and novel approach that integrates preference learning with region-specific optimization, offering a meaningful advance for clinically grounded medical image generation.
- 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 claim of alignment with ‘clinical preferences’ is supported only by quantitative metrics like PSNR, SSIM, LPIPS. Without evaluation by clinical experts, it is unclear whether the generated images are truly realistic or clinically useful. This limitation should be addressed or acknowledged. Glioblastomas are inherently 3D, with complex volumetric structure and infiltration patterns. Modeling generation in 2D discards important spatial context, raising concerns about clinical validity. The authors should justify this design choice or clearly acknowledge it as a limitation. Longitudinal image generation has been explored in prior studies, including works using publicly available datasets such as LUMIERE. The manuscript would benefit from a clearer discussion of how the proposed approach differs from and improves upon existing methods applied to similar longitudinal datasets.
- 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 proposed tumor-aware DPO framework represents a meaningful extension of preference-based learning, specifically designed to capture clinically relevant structures. The method produces more clinically plausible outputs than prior generative approaches, particularly in modeling tumor regions and their progression over time.
- 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
Author Feedback
N/A
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 proposes Tumor-Aware Direct Preference Optimization for longitudinal post-operative MRI generation. All three reviewers agree that the problem is clinically meaningful and technically challenging. There are some minor concerns from the reviewers. Reviewer 1 notes that the dataset is split at the slice level rather than the patient level, which may introduce information leakage. Reviewer 2 questions whether the notion of “preference” is truly clinical preference or simply ground-truth-guided alignment, and notes the lack of comparison with other preference optimization or RL-based methods. Reviewer 3 emphasizes the need for clinical expert evaluation of generated images, since the current evaluation relies only on PSNR, SSIM, and LPIPS without clinical validation.