List of Papers Browse by Subject Areas Author List
Abstract
Breast reconstruction plays a crucial role in restoring body image and psychosocial well-being after mastectomy, yet preoperative planning relies heavily on surgeons’ experience and lacks practical planning tools. In this work, we present a controllable generative framework to simulate postoperative breast reconstruction outcomes from preoperative chest CT. Our pipeline comprises three components: (1) a conditional 3D latent diffusion model with an inpainting scheme that enables high-fidelity synthesis within the surgical region; (2) an implant canonical template library and anatomy-aware alignment method to construct a physiologically plausible implant spatial prior for guiding the generation; (3) a dedicated control branch that injects the implant prior into the diffusion process to enforce the implant and breast geometry changes. Experiments show that our framework can generate realistic postoperative volumes, and spatial prior injection substantially improves the geometry accuracy, leading to superior performance over competing baselines in both fidelity and controllability. We further demonstrate that simulated outcomes vary coherently with the user-defined implant configurations. Overall, our work enables preoperative visualization under diverse surgical configurations and supports decision-making with a flexible, patient-specific planning tool for breast reconstruction. The source code is available at https://github.com/wangyz0915/DiffuBreast.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4830_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/wangyz0915/DiffuBreast/
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WanYiz_Preoperative_MICCAI2026,
author = { Wang, Yizhi AND Yu, Tianze AND Yu, Hengjie AND Liu, Maoling AND Pan, Wenxuan AND Ding, Zhongxiang AND Fu, Peifen AND Jin, Yaochu},
title = { { Preoperative Simulation of Personalized Breast Reconstruction via Implant Prior Controllable Diffusion } },
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
This paper proposes a controllable 3D generative framework for simulating postoperative outcomes of breast reconstruction using preoperative chest CT scans. The method utilizes a 3D Latent Diffusion Model (LDM) adapted for an inpainting task within a surgical Region of Interest (ROI). To ensure anatomical and geometric fidelity, the authors introduce a “spatial implant prior” constructed from a canonical template library. This prior is aligned to the patient’s ribcage anatomy and injected into the diffusion process via a ControlNet-like branch. The framework is evaluated on a retrospective dataset of 664 subjects, comparing against various conditional diffusion baselines and spatial control variants. Results indicate improved structural similarity and implant placement accuracy compared to existing methods
- 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 choice of an inpainting LDM is technically sound for this task, as it maintains the integrity of the contralateral breast and non-surgical anatomy while focusing generative capacity on the ROI.
2.The authors correctly identify that implicit conditioning (cross-attention on text/numeric attributes) is insufficient for enforcing strict 3D geometric constraints in medical imaging, justifying the need for explicit spatial control.
3.The use of an external test set from a different hospital helps assess the model’s generalization capabilities across different scanning protocols.
- 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 novelty: The core architecture is essentially a 3D adaptation of ControlNet applied to a 3D LDM inpainting framework. While the application is novel, the methodological components are standard modules in the current generative AI landscape.
2.The comparative experiments are insufficient. The comparative methods are all traditional generative methods from several years ago; it is necessary to include newer methods from the last 2-3 years. In addtion, GLIGEN and T2I-Adapter are 2D models, I am concerned about the fairness of these comparative experiments.
3.You mention excluding cases with “severe implant deformation”. Does this imply the model is only trained on “ideal” outcomes?
4.In Section 2.4, the intermediate denoised estimation is passed through a frozen VAE decoder and a segmentation network. As LDMs are known to produce noisy intermediate approximations at earlier timesteps, it is questionable how robust the frozen segmentation network is to these out-of-distribution, noisy predictions. The authors do not state if the SwinUNETR was trained with noise augmentation.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 primary factor for my recommendation is the transition from implicit to explicit spatial conditioning. While the algorithmic novelty is somewhat incremental (recombining 3D LDM, ControlNet, and inpainting), the specific design choices, such as the ribcage-based alignment and the use of canonical templates, are well-justified for the medical domain. The experimental validation on a large, multi-center dataset is a significant strength.
- 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 author has addressed most of the issues I raised, and the overall quality of the manuscript warrants consideration for acceptance by miccai2026.
Review #2
- Please describe the contribution of the paper
The paper presents a dedicated 3D generative pipeline for post-mastectomy reconstruction simulation that outperforms general-purpose controllable diffusion models. Extensive experiments and visualizations validate the proposed pipeline’s effectiveness and usability.
- 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 work addresses a significant gap in surgical planning by moving beyond qualitative visualization toward reliable, patient-specific 3D morphological simulation.
- The framework allows for the adjustment of height, width, and projection, enabling surgeons to visualize outcomes for implants with varying properties
- 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.
- While understandable, the paper uses the mask of the implant (I_p) as a part of the model’s input, which may hinder its downstream application, since the accurate geometry of the implant itself is usually hard to get before the surgurical operation.
- This paper uses Totalsegmentator for ribcage segmentation and MedVAE for LDM autoencoding, so the actual performance of breast reconstruction is also dependent on the pretrained multi-organ model/autoencoder themselves. – Following the above weakness, if the authors opt to use pretrained models for medical priors, why not directly use diffusion networks pretrained on medical data with additional adapters to realize task-specific synthesis. – Current comparisons (GLIGEN, Siimple-Controlnet, T2I-Adapter) are not tested or pretrained on medical data. Also, they were designed as 2D models, it is important that the authors discuss the reasons how they elevate them to 3D CT synthesis as well as why not choose 3D synthesis methods in the medical domain as baselines.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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.
(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?
1.Require mask input, which from my perspective is often hard to get for downstream applications of interest. 2.Comparisons are not revolved in the medical domain, there are currently many diffusion-based controllable pipelines for medical CT synthesis (e.g. GuideGen, MedSyn, MAISI, …)
- 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.
I appreciate the authors’ rebuttal effort, however, I am not satisfied with the current version. I think authors can benefit more from comparing with more recent inpainting methods can demonstrate their ineffectiveness through experiments, and implement novelties based on them.
Review #3
- Please describe the contribution of the paper
This paper presents a controllable, patient-specific generative framework for simulating postoperative breast reconstruction outcomes from preoperative chest CT scans. Its main contribution lies in integrating a 3D latent diffusion model for inpainting with an explicit, anatomically aligned implant spatial prior, constructed from a library of canonical templates and adapted to each patient’s anatomy. Injecting this prior into the model via a dedicated control branch enables precise control over implant geometry and placement while preserving the consistency of surrounding tissues, resulting in realistic and anatomically coherent postoperative simulations. Compared to prior work, the approach stands out for explicitly modeling surgical variables in three-dimensional space and demonstrating significant improvements in anatomical fidelity and implant geometry accuracy. It therefore provides a tool for preoperative planning, allowing controlled and personalized exploration of different implant configurations.
- 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.
This paper introduces an approach for controllable postoperative simulation by explicitly integrating a patient-specific 3D implant spatial prior within a latent diffusion framework and enables direct geometric control over surgical outcomes. The use of a canonical implant template library, combined with anatomy-aware alignment, particularly via principal component analysis (PCA) of the ribcage, provides a way to translate discrete implant specifications into a continuous and physiologically plausible spatial representation. The method is closely tailored to the clinical problem through the use of ROI-based inpainting. This mechanism allows realistic modification of only the surgical region while preserving surrounding anatomy, which is crucial in unilateral reconstruction scenarios. The work demonstrates clear clinical utility by enabling simulation of postoperative outcomes under different user-defined implant configurations. It thus offers a practical tool for preoperative planning and surgical decision-making. Finally, the experimental evaluation combines standard image quality metrics with implant-specific geometric measures (such as Dice and IoU), along with qualitative analyses. The inclusion of cross-institutional testing further strengthens the claims regarding the robustness and generalizability of the proposed method
- 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 also presents several important weaknesses. First, the methodological novelty may appear relatively incremental, as the approach largely builds on existing components such as latent diffusion inpainting (e.g., RePaint), spatial conditioning via ControlNet-like architectures, and template-based anatomical priors without introducing a fundamentally new generative paradigm. Prior work has already explored spatially guided medical image synthesis, particularly in the context of controllable diffusion and anatomy-constrained generation. Moreover, although the construction of the implant prior is well motivated, it primarily relies on aggregating aligned segmentations and on PCA-based alignment. This strategy may lead to an oversimplification of the complex biomechanical interactions between the implant and surrounding soft tissues, and does not explicitly model tissue deformation. In this respect, it remains less expressive than physics-based or biomechanical simulation approaches commonly used in surgical planning. Additionally, the clinical applicability of the method remains limited. The approach assumes the availability of preoperative CT scans, which are not systematically acquired in all breast reconstruction workflows. Moreover, it focuses exclusively on implant-based reconstruction and excludes more complex cases, such as those involving severe deformation or displacement. These choices raise concerns regarding robustness in real-world settings and potential selection bias. Furthermore, comparisons are restricted to general-purpose controllable diffusion methods adapted to the medical setting, rather than to specialized surgical simulation tools. Finally, the claim that the method enables exploration beyond manufacturer implant catalogs through continuous scaling is not clinically validated. This raises questions about the anatomical plausibility of the generated results and their relevance within real-world regulatory constraints.
- 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.
(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?
based on the major weaknesses mentioned
- Reviewer confidence
Very confident (4)
- [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.
N/A
Author Feedback
We appreciate R2 for recognizing the clinical value and multi-center validation, R6 for the detailed technical scrutiny that improved our methodological clarity, and R8 for the profound clinical insights that for sure benefit our future work. We address concerns below. 1) Baseline selection R2W2 R6W4: Fairness is ensured: All baselines were elevated to 3D by replacing 2D conv/attention modules with 3D counterparts, sharing the same 3D LDM backbone, training data, and training iterations. Their reasonable performance (Fig. 2, Tables 1-2) supports the correct adaptation. We will release their code as well, and elaborate the adaptation method in the camera-ready version. We chose baselines to cover representative and recent control methods: GLIGEN2023, T2IAdapter2024, SimpleControlnet2025.Medical models like GuideGen, MedSyn, MAISI target different tasks rather than breast reconstruction, causing domain mismatches and thus are not directly comparable. 2) Novelty R2W1 R8W1: We clarify that our novelty is at the system level for an underexplored clinical task: (i) LDM inpainting for simulation; (ii) template library bridging historical postoperative knowledge to preoperative use; (iii) continuous shape control via template editing; (iv) PCA alignment for anatomical plausibility; (v) control branch improves implant geometry accuracy and sharpness. While (i) follows LDM, all components are task-oriented designs. These formulate, to our knowledge, the first open-source framework for patient-specific breast reconstruction simulation, with continuously adjustable implants. 3) Implant mask R6W1: We agree that the postoperative implant is unknown preoperatively, which motivates Sec.2.3.Importantly, I_p is NOT the postoperative implant of the target case but a planned control signal obtained by template retrieval, size/placement adjustment, and PCA anatomy alignment; thus, it is available preoperatively. Deformations: The same type/size implants show similar mild bottom bending due to skin coverage and rib support. We approximate this by averaging historical implants of the same type/size (red/blue overlap in Figs. 2 3 supports this). 4) Model dependency R6W2: While we use large-FoV CTs for body visualization, we thank the reviewer for noting the potential failure. Quantification: Monitoring rib segmentation volume and HU consistency of bone; Mitigation: if it fails, fall back to bone window thresholding, filtering, connectivity based rib estimation within the ROI as an alternative. 5) Use of pretrain R6W3: As baselines: domain mismatch (see 1). As generation backbones: they are pretrained on general CTs with few implants, lacking implant prior. Injecting control may not reuse their knowledge well and overload the control branch for implant generation, causing instability. Therefore, we train the backbone on domain data. 6) Plausibility of deformations R8W2: Targeting the simulation with accurate shape prior, we did not explicitly model biomechanics interaction. We will add it as limitation and explore physics-constrained generation in future work. 7) Clinical use R8W2: CT availability: Chest CT is acquired for mastectomy and reconstruction for anesthesia planning and lung, lymph metastasis assessment. We appreciate the R8’s practical concern and will explore modality transfer to enhance data availability. Clinical validation: Breast simulators are commercial (for augmentation) or photo-based non public models, hindering direct comparison. Prospective validation needs a longer timeline and will be pursued. Additionally: Outlier cases R2W3 R8W2: Our model aims to simulate the intended postoperative breast shape; rare abnormal cases were excluded to avoid ambiguous generation. Noisy prediction R2W4: SwinUNETR receives denoised intermediate approximation (Eq.4), which is close to a clean image, not noisy; so noise augmentation is unnecessary. We sincerely thank reviewers for their constructive comments, and invaluable feedback.
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.
The use of generative methods for predicting surgical outcomes using patient-specific anatomical constraints and possible implants is a sound concept. Nevertheless the reviewers raised important issues that need to be addressed by the authors. It is requested to the authors to explicitly explain how these aspects can be remedied within the rebuttal phase.
In concrete the following aspects need to be discussed:
1) Baseline selection: The major issue detected by the reviewer was the unfair comparison to the baselines (R2, W2; R6, W4) - please explain why this is not the case, or how to remedy this. 2) Novelty: Reviewers claim all components are known and the novelty comes from the integration (R2, W1; R8, W1) - please comment on that. 3) Implant mask: Implants are deformable on their own, such that the implant mask cannot be known before surgery (R6, W1) - clarify if that statement is true and if yes, how this impacts predictions. 4) Dependency on pretrained models: Frozen models can produce errors that propagate downstream (R6, W2) - what measures have authors taken to quantify and mitigate this, in the view that TotalSegmentator does not perform well on ribs if CTs have a small FoV. 5) Use of diffusion networks: Diffusion networks pretrained on medical data exist (R6, W3) - explain the rationale of not using them at least as baseline. 6) Plausibility of deformations: The approach does not guarantee realistic biomechanical deformations (R8, W2) - elaborate if evaluation of deformation fields was performed and at least mention this in the limitation section 7) Clinical use: Limitations of an introduction of the method in a clinical setup are not addressed (R8, W2) - discuss them.
Notation: Rx stands for reviewer x; Wx stands for weakness x (being x the paragraph of answer 7)
- 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.
While I do not think that the authors addressed all raised issues, the main criticism (the fairness of comparison with the baselines which would have been a “deal-breaker”) seems to be solved. While the two reviewers giving an “Accept” remain short in justification, the justification of reviewer 6 to reject the paper is not really valid in my view as it misses the point of the approach: other inpainting methods could be replaced in the same architecture and the paper would still be novel.
I still hope the authors keep their word an implement the changes they promise in the camera-ready version. I still believe however they should not claim that deformation is “mild”. That is definitely not the case after working >15 in breast cancer. I recommend addressing this as a limitation: “the model will perform fine as long as the implants do not deform much”. Also, the rebuttal answer for 4) is weak… Again here I recommend explicitly mention this limitation and possibly pointing at literature “correcting” total segmentator or alike. The proposed strategy is also OK, but without implementing it, it should at most be considered a potential future work. The issue with 7) is not to be dismissed that easily. I agree with the line of thoughts of the authors (cases where CT is available and potentially using othermodalities), but porting their model to MRI or US is not going to be easy.
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.
The rebuttal is satisfactory
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.
This work addresses an important surgical planning problem and proposes a controllable 3D generative framework for personalized breast reconstruction simulation using an anatomically aligned implant spatial prior. The explicit implant prior, ROI-based inpainting formulation, and cross-institutional evaluation provide clinically meaningful strengths, and the rebuttal satisfied most concerns for two reviewers. The remaining concerns about incremental novelty, comparison with more recent medical-domain inpainting or controllable diffusion methods, reliance on implant priors, and limited modeling of soft-tissue deformation should be addressed in the final version.
