List of Papers Browse by Subject Areas Author List
Abstract
Breast-conserving surgery (BCS) is the preferred treatment for early-stage breast cancer, offering survival rates comparable to mastectomy while preserving breast aesthetics. Accurate tumor segmentation is essential for surgical planning, yet segmentation models often exhibit biases toward specific tumor sizes, particularly underperforming on smaller tumors. To address this, we propose a novel approach that uses generative models to improve segmentation across tumor sizes. Specifically, we adapt the Stable Diffusion model and apply a Denoising Diffusion Probabilistic Model (DDPM) inversion approach to generate synthetic tumors of controlled sizes within real breast MRIs, helping to balance tumor size distribution in the training data. By augmenting the dataset with 10–20% synthetic tumor images, our method significantly improves segmentation accuracy for small tumors without compromising performance for larger tumors. This enhancement allows for more precise tumor assessment, leading to better-informed surgical decisions and potentially reducing unnecessary mastectomies.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/5247_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{LunMig_Improved_MICCAI2025,
author = { Luna, Miguel and Baek, John and Kim, Won Hwa and Son, Wan Gyu and Lee, Kwang Min and Kim, Hye Jung and Kim, Jaeil},
title = { { Improved Tumor Segmentation using Selective Synthetic Augmentation for Enhanced Surgical Planning in Breast MRI } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15970},
month = {September},
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces a framework using Stable Diffusion and DDPM for generation of size-controlled synthetic tumors in breast MRI images and evaluates segementation performance of their segmentation model (SwinUNetr) across tumors of varying sizes. The segmentation model is trained on a mix of real and synthetic images and then fine-tuned on only real data. The authors show that their method improves the segmentation performance, specifically across small tumor category, when using 10-20% synthetic data in addition to real MRI images.
- 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 motivation of the task is very clear and well-stated. 2. The idea could be very useful in real case scenaios where obtaining annotations for medical images is very time-consuming and laborous effort. 3. The method is novel to some extent, specially for breast MRI-representing a promising direction for controllable synthetic data generation for augmenting data. 4. Use of LoRA for fine-tuning with a focus on reduced computational cost is a good effort. 4.I also found the size-controlled tumor generation is an interesting aspect, which could possibly help with segemntation performance bias in small tumors. Fine-tuning with size-specific prompts shows a methodical attempt to align synthetic tumors with desired clinical dimensions.
- 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 clinical validation for the synthetic data is a big hindrance in application of this solution. There is no mention of expert radiologist involvement in assessing the realism of synthetic tumors. 2. The mention of using “prompt strength” to control realism is interesting, but it feels subjective. There is no way to quantify how realism drops off with higher prompt strengths, for e.g., any reader studies or perceptual evaluations. 3. DDPM inversion’s stochastic nature raises concerns about consistency and anatomical plausibility across repeated runs. 4. All data generation and insertion are done slice-wise in 2D, ignoring 3D tumor continuity, which is vital for clinical applications like surgery and treatment planning. 5. The segmentation performance of the model is evaluated only based on Dice scores, which seems incomplete without the inclusion of atleast a distance metric. Spatial accuracy is as important of a factor as overlap is. Two tumor predictions with same Dice score might look very different on the MRI image to a radiologist.
Minor: One paragraph in Introduction section on page to is repeated with slight difference in phrasing, both starting with “Generative models can…”
- 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.
- 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
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.
(3) Weak Reject — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Lack of clinical validation of the generated synthetic tumor data is a big factor in the way of its acceptance. There is no way to ensure that the generated tumors actually resemble real tumor morphology seen in breast MRI. The lack of proper evaluation metrics further undermines the credibility of the results.
- 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 is well-written and presents an intriguing approach to addressing an important clinical problem. It focuses on segmenting relatively small tumors (<2cm) in MRI scans. A latent diffusion model is employed to generate synthetic tumors, which are then integrated into real MRIs using a transitional diffusion model. Segmentation results are reported based on training with varying percentages of synthetic data.
- 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 research statement is well-motivated, with a clear articulation of the clinical problem.
- The concept of generating controlled-size tumors within real MRI 2D slices using diffusion models is a plausible approach.
- Combining latent and transitional diffusion models to enhance image segmentation is a sensible strategy, balancing both time and accuracy.
- Focusing on lumpectomy offers a clinically relevant and feasible solution.
- 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.
- Mismatch between Figure 3 and Table 1 results, suggesting data imbalance; standard deviations, violin plots, and segmentation mask examples are needed for better validation.
- Tumor morphology, especially irregular versus smooth shapes, was not adequately considered, limiting the robustness of the evaluation.
- Clinical relevance would have been stronger if BI-RADS scoring was incorporated into the assessment framework.
- 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
Detailed comments:
-
Results Reliability: There appears to be a mismatch between the results presented in Figure 3 and those reported in Table 1. While Table 1 shows mean Dice scores around 80%, Figure 3 does not seem to support these values, suggesting a potential inconsistency. This may be due to differences between the mean (reported in Table 1) and the median (depicted in the box plots), which would indicate remaining data imbalance. The authors are encouraged to justify this observation, report standard deviations alongside mean scores, and consider using violin plots to more clearly represent the full distribution and density of the Dice scores, particularly given the skewness visible in the current box plots.
-
Tumor Morphology Evaluation: Although the work addresses tumor size variability, it does not adequately consider tumor morphology, which is a critical factor in breast cancer diagnosis. The authors have acknowledged limitations regarding unifocal tumors, but incorporating irregular versus smooth tumor shapes during the synthetic tumor map design would have added valuable depth and robustness to the evaluation. This enhancement could have been straightforward to implement and would further strengthen the findings.
-
Illustration Requirements: To support the reported quantitative Dice scores, segmentation masks with ground-truth overlaps should be shown. Visual examples are essential to validate the quality of the segmentation and to provide a more intuitive understanding of performance.
-
Clinical Relevance Consideration: For true clinical applicability, the evaluation should have incorporated BI-RADS scoring, as radiologists base decisions on this standardized assessment rather than on subjective impressions. While not expected to be added at this stage, it should have been considered during the initial methodological design.
-
Writing Improvement: The manuscript contains redundancy, particularly between the second and third stated contributions, which could be merged for clarity. Similarly, the first paragraph of Section 2.4 repeats ideas already discussed, indicating a need for tighter and more concise writing.
-
- 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?
The paper is well-written and addresses a critical clinical issue: The approach using diffusion models for tumor generation and segmentation is plausible. The focus on lumpectomy offers a clinically relevant solution.
- 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 authors present a novel approach for improving breast tumor segmentation by using text-guided diffusion models to selectively augment the training data with synthetic tumors of specific sizes. By leveraging Stable Diffusion with DDPM inversion, they generate realistic synthetic breast MRI scans conditioned on tumor size, thereby addressing a key limitation in segmentation models: underperformance on small tumors. The method achieves improved segmentation accuracy on small tumors (<2cm) without compromising performance on larger lesions, offering a practical tool for enhanced surgical planning in breast-conserving therapy.
- 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.
- Tackles a clinically meaningful problem with direct relevance to surgical decision-making in breast cancer.
- Technically innovative use of DDPM inversion and text-guided control for synthetic tumor generation
- Uses synthetic augmentation strategically and selectively (10–20%), avoiding common pitfalls of over-synthesis or domain shift.
- Demonstrates impact across tumor sizes, with stratified results for <2cm, 2–5cm, and >5cm tumors, which is thoughtful and clinically aware evaluation.
- Clear link between the proposed method and downstream decision support (i.e., improving BCS planning).
- 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.
- Only Dice scores are reported. The inclusion of boundary-aware metrics such as Hausdorff Distance or MSD would have provided a more complete view of segmentation quality, especially in small tumor cases.
- No statistical significance or variance measures are reported, which limits confidence in the robustness of the observed improvements.
- 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
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.
(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?
This paper presents a technically creative and clinically valuable method for improving tumor segmentation via controlled synthetic augmentation. The thoughtful use of text-prompted diffusion models, size-stratified analysis, and practical augmentation ratios suggest a high level of rigor and awareness. While the lack of distance metrics and variance reporting are limitations, they are relatively minor and can be easily addressed.
- 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
We sincerely thank the Reviewers for their thoughtful and constructive feedback. We are encouraged by the recognition of our work’s clinical relevance, innovation, and practicality. Below, we address specific points raised during the review process.
Boundary-Aware Metrics We appreciate the suggestion to include distance-based metrics to complement the Dice score. While we are unable to include new results at this stage, we agree that such metrics could provide valuable insights. However, for small lesions where overlapping errors may be more pronounced, improved Dice scores may also indirectly indicate reduced boundary errors.
Results Presentation We acknowledge that the presentation of results can be further enhanced. Including statistical significance testing, variance measures, and additional visualizations such as overlay images could improve interpretability. Where space permits, we intend to incorporate these enhancements in the final version as they serve to strengthen the current findings.
Mean and Median Dice Scores Reviewer #1’s observation is correct: Table 1 reports the mean Dice score across different lesion size subgroups, while Figure 3 presents the median Dice score (represented by the center line in the boxplot) across the full dataset. We will clarify this distinction in the final manuscript to avoid potential confusion.
BI-RADS Categories and Lesion Shape We agree that subgroup analyses by BI-RADS category and lesion shape (e.g., oval, round, irregular) are clinically meaningful. Although new analyses cannot be introduced post-submission, we consider these important directions for future work.
3D Continuity We agree that preserving 3D continuity is important, particularly for clinical applications such as treatment planning. Due to practical considerations with stable diffusion, our pipeline was developed in 2D. Evaluation, however, was performed on 3D volumes to ensure spatial coherence and clinical relevance. While the current approach does not explicitly enforce 3D consistency, incorporating mechanisms that can provide 3D context to 2D models represents a promising direction for future work.
Anatomical Consistency and Stochasticity We acknowledge the potential for anatomical variation due to the stochastic nature of DDPM inversion. However, synthetic images were used solely for representation learning, and fine-tuning was performed exclusively on real images, ensuring anatomically grounded predictions. The stochastic process also enhances sample diversity, which benefits robust feature learning.
Tumor Realism and Prompt Strength As shown in Figure 1b, prompt strength significantly influences the realism of generated tumors. Higher values may reduce realism, while lower values may result in limited changes. We used expert feedback to heuristically select a prompt strength that balances realism and the intended modifications. While this input was not explicitly stated in the original paper, we will clarify it in the final version.
Redundancy in Text We agree that the manuscript can be improved by removing redundant phrasing. We will revise the text for greater clarity and conciseness.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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