List of Papers Browse by Subject Areas Author List
Abstract
Computational phantoms are widely used in medical imaging research, yet current systems to generate controlled, clinically meaningful anatomical variations remain limited. We present AbdomenGen, a sequential volume-conditioned diffusion framework for controllable abdominal anatomy generation. We introduce the Volume Control Scalar (VCS), a standardized residual that decouples organ size from body habitus, enabling interpretable volume modulation. Organ masks are synthesized sequentially, conditioning on the body mask and previously generated structures to preserve global anatomical coherence while supporting independent, multi-organ control. Across 11 abdominal organs, the framework generates realistic anatomies, achieving highest fidelity for major solid organs (e.g., liver Dice 0.83 ± 0.05). VCS yields stable single-organ calibration over [−3, +3] and disentangled multi-organ modulation. To showcase clinical utility on a hepatomegaly cohort selected from Merlin Dataset, Wasserstein-based VCS selection reduces distributional distance by 73.6% relative to the training distribution. These results demonstrate calibrated, distribution-aware anatomical generation suitable for controllable abdominal phantom construction.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5543_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{BhaYub_AbdomenGen_MICCAI2026,
author = { Bhandari, Yubraj AND Dahal, Lavsen AND Segars, Paul AND Lo, Joseph Y.},
title = { { AbdomenGen: Sequential Volume-Conditioned Diffusion Framework for Abdominal Anatomy Generation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16889},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents a diffusion-based framework for generating 3D shapes across 11 abdominal structures. Specifically, habitus-decoupled conditioning variables are introduced to resolve the contradictory-signal problem of raw-volume conditioning. A sequential autoregressive generation strategy is utilised to ensure the spatial coherence and independent organ-wise control.
- 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) This is a clinically well-motivated study. (2) The experiment design is comprehensive and strong.
- 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 evaluation only covers 4 of 11 organs in the main Table. (2) Lack of justification on the order of sequential generation. (3) The results of the Kidney are poor and unexplained.
- 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
(1) It is suggested to include the results of the other seven organs in Table 1. (2) Future works can explore the ablation of different organ generation orders. (3) What is the inference time of the whole generation process? (4) The habitus-decoupling model assumes a linear relationship between body volume and organ volume. However, for organs such as the gallbladder and stomach (r = 0.28), this assumption is likely violated. (5) Please clarify the reason for setting the size parameter of 0.9 for MAISI?
- 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?
Overall, this paper is good. It addressed a clinically meaningful problem and designed a strong experiment to validate their methods.
- 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.
The authors have successfully addressed most of my concerns. It is highly recommended to include more evaluation and develop a fast version of the generation in the Journal version of this paper.
Review #2
- Please describe the contribution of the paper
This paper proposes AbdomenGen, a sequential diffusion framework for generating abdominal organ masks with organ-specific volume control. The key idea is the Volume Control Scalar (VCS), a standardized residual that aims to separate organ size from overall body habitus. The model synthesizes organs sequentially, conditioning on the body mask plus all previously generated structures. The authors conduct experiments on 556 abdominal CT scans, reporting on shape fidelity, diversity, single-organ calibration, limited multi-organ disentanglement, and a liver-focused distribution-matching case study for hepatomegaly.
- 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.
Important and Underexplored Problem: Controllable generation of anatomically coherent abdominal phantoms is genuinely useful for simulation, stress testing, and dataset enrichment. The focus on multi-organ controllability rather than solely image realism is well-motivated, particularly the distinction between fidelity, diversity, and clinically meaningful control.
Sensible Core Design: The sequential generation strategy is intuitive for the abdomen, where organs have strong geometric dependencies and compete for space. The use of a body SDF plus an accumulated context SDF is a highly practical way to encode global occupancy without drastically increasing the input dimensionality as the organ count grows.
Interpretable and Clinically Legible VCS: The formulation in $Equation (1)$ yields a simple, standardized residual relative to predicted organ volume from body volume. It is an easy-to-understand and highly usable metric, arguably preferable to raw latent space traversals for controlled phantom construction. The authors successfully motivate why raw organ volume can be contradictory when body habitus already provides correlated information.
Clear Model Description: Figure 1 does an excellent job explaining the sequential pipeline, clarifying how the generated SDF feeds back into the evolving context and how FiLM injects the scalar control.Strong Evidence of Controllability: Figure 3(a) is a standout, demonstrating monotonic volume response across multiple organs over the VCS sweep. The approximately linear trend for the liver and kidneys supports the claim that the control variable is properly calibrated.
Interesting External-Cohort Matching: The hepatomegaly experiment is practically relevant. Figure 3(b) and the reported Wasserstein reduction successfully indicate that the method can shift a generated population toward a target pathology-related volume regime without retraining.
- 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.
Narrow Empirical Validation Relative to Claims: The paper repeatedly frames the contribution as controllable generation over 11 abdominal organs. However, the quantitative evaluation (Table 1) reports only four groups (liver, spleen, kidneys, stomach), and Table 2 is liver-only. Qualitative evidence (Figure 2) and joint-control evidence (Figure 4) focus exclusively on the liver and spleen. The main paper lacks quantitative metrics for the colon, rectum, duodenum, pancreas, urinary bladder, or small bowel. Omitting the most difficult, compact, or hollow organs undermines the broad multi-organ claims.
Weak Baseline Comparison: The comparative result in Table 2 benchmarks AbdomenGen against MAISI only on liver diversity, using restricted metrics (pairwise Dice and Chamfer) under a fixed MAISI size parameter. There is no head-to-head comparison for fidelity, calibration accuracy, multi-organ independence, or external distribution matching, which does not support the claim that “no existing method satisfies all three requirements.”
Lack of Direct VCS Ablations: The central novelty is the habitus-decoupled scalar, but there is no direct ablation against conditioning on raw organ volume, log-volume, percentile-normalized volume, or a nonlinear body-volume predictor. It is not convincingly demonstrated that the residualization step is practically necessary rather than just conceptually tidy.
Simplistic Linear Habitus Model: Expected organ volume is modeled as a linear function of body volume, yet reported correlations vary widely (liver at 0.64 down to spleen/stomach at 0.24). For weakly correlated organs, a linear residual may be a noisy normalization.
Mixed Fidelity Metrics: Table 1(A) shows highly unstable numbers for kidneys (Dice 0.60 ± 0.23, HD95 16.05 ± 22.50). This spread indicates real failure cases or strong instability, contradicting the text’s assertion that this merely reflects localized boundary deviations.
Inconclusive Multi-Organ Independence: Figure 4 suggests axis-aligned shifts for the liver and spleen, but this is an isolated two-organ, two-dimensional case study. There is no quantitative independence metric, no confidence intervals, and no analysis of cross-organ volume drift in nearby organs.
Incomplete Realism Argument: The manifold realism analysis via nearest-neighbor distance (Table 1) is a good sanity check against memorization, but it is not strong evidence of sample realism. Nearest-neighbor surface distances can easily miss topology issues in elongated/tubular structures.
Volume-Heavy External Experiment: The 73.6% reduction in Wasserstein distance is optimized for 1D liver volume. Figure 3(b) demonstrates scalar distribution alignment, but does not prove anatomically faithful hepatomegaly morphology, as Table 1 is not stratified by the target regime.
Missing Related Work: The paper fails to discuss several highly relevant works addressing controllable abdominal and full-torso generation:
Dai et al., “GuideGen: A Text-Guided Framework for Full-torso Anatomy and CT Volume Generation” (2024).
Konz et al., “Anatomically-Controllable Medical Image Generation with Segmentation-Guided Diffusion Models” (2024).
Wang et al., “Devil is in Details: Locality-Aware 3D Abdominal CT Volume Generation for Organ Segmentation” (2024).
Zhuang et al., “Semantic Image Synthesis for Abdominal CT” (2023).
Belton et al., “DyABD: A Dataset and Technique for Synthetically Generating Dynamic Abdominal MRIs…” (2024).
- 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
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?
The paper tackles a highly relevant application with a sensible core idea (sequential diffusion with VCS conditioning). However, the current evidence is simply too partial for the broad scope of the claims made. The major factors driving my “Weak Reject” recommendation are the exceptionally limited baseline comparisons, the lack of direct ablations validating the necessity of VCS over simpler alternatives, and the incomplete validation across the advertised 11 organs (with several of the most challenging organs entirely omitted from the quantitative results).
My score could increase if the authors convincingly show stronger all-organ evaluation and direct ablations establishing the benefits of VCS during the rebuttal. Conversely, my score would likely decrease to a Strong Reject if the omitted organs perform poorly or if VCS offers little empirical benefit over raw-volume conditioning.
- 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.
After reading the rebuttal, I am willing to raise my recommendation to Accept, although still with some reservations.
The rebuttal addresses my main concerns reasonably well. In particular, the authors provide additional all-organ nearest-neighbor analysis, clarify the poor kidney and colon results, and give a direct comparison between raw volume, z-score volume, and VCS conditioning. The VCS ablation is especially helpful and makes the central design choice more convincing. The authors also clarify the MAISI setting, inference time, and the relation to recent controllable generation work.
Some concerns remain. The strongest evidence is still concentrated on volume-level control rather than richer anatomical morphology; kidney and colon performance remains weak; and the comparison to existing controllable generation methods is still limited. Also, several important analyses are only provided in the rebuttal rather than in the submitted manuscript. Nevertheless, I think the paper addresses a relevant and underexplored problem, proposes a sensible and interpretable control formulation, and provides enough evidence after rebuttal to be marginally above the acceptance threshold.
Review #3
- Please describe the contribution of the paper
The paper proposes a diffusion-based framework for generating multi-organ abdominal anatomies with explicit control over organ size. The key idea is the introduction of a Volume Control Scalar (VCS), which models organ volume as a residual with respect to body size and serves as a control variable during generation.
The authors also adopt a sequential generation strategy, where organs are generated one at a time conditioned on the body mask and previously generated structures.
The framework is further used to match target population distributions (e.g., hepatomegaly) by adjusting the VCS without retraining.
- 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.
One strength of the paper is the formulation of the Volume Control Scalar (VCS). Normalizing organ volume with respect to body size leads to a more interpretable and decoupled control variable, which addresses the well-known issue that raw volume is entangled with global body habitus.
The sequential generation strategy also makes sense in this context. Generating organs one at a time while conditioning on previously generated structures helps maintain spatial consistency and makes per-organ control easier compared to joint generation.
The experimental evaluation is fairly comprehensive. Beyond standard geometric metrics, the paper looks at controllability, multi-organ independence, and distribution matching. The use of Wasserstein distance to match a target population distribution is a nice touch and shows a potential use case beyond single-sample 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.
While the idea of introducing VCS is reasonable, the overall methodological novelty feels somewhat limited. The framework mainly builds on standard components (diffusion models, FiLM-style conditioning, and sequential generation), and comes across more as a combination of existing ideas rather than a fundamentally new formulation. It would help to position the method more clearly with respect to prior work on controllable generative models, especially those using latent or scalar-based controls.
The sequential generation strategy is intuitive, but it may introduce error accumulation across organs. This is briefly mentioned, but not really analyzed. It would be useful to see an ablation or some quantitative study on how sensitive later organs are to earlier predictions.
The evaluation covers several aspects, but is still largely limited to controlled settings. In particular, the controllability experiments (e.g., VCS sweeps and hepatomegaly matching) rely on simulated manipulations, and there is no validation on downstream tasks or clinical workflows. This makes it hard to judge the practical utility of the approach.
The experimental comparison is also somewhat limited. The comparison with MAISI mainly focuses on diversity, and it is not clear how the method performs against other recent generative or controllable approaches.
Finally, the model only controls volume, which is a fairly coarse descriptor of anatomy. Volume changes do not necessarily reflect clinically relevant variations in shape or local morphology, and the current formulation does not address this.
- 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.
(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?
I think the paper addresses an interesting problem and presents a reasonably well-executed solution. The idea of introducing a normalized control variable for organ volume is intuitive, and the experiments are fairly comprehensive in evaluating controllability and distribution matching.
That said, I am not fully convinced about the level of novelty. While the VCS formulation is a reasonable design choice, the overall framework largely builds on existing components (diffusion models, FiLM-style conditioning, and sequential generation), and the contribution feels somewhat incremental.
In addition, the practical relevance is somewhat limited at this stage. While the paper demonstrates controllability through VCS sweeps and distribution matching, there is no evaluation on downstream tasks or real clinical workflows, making it difficult to assess the actual utility of the proposed framework.
Overall, while the paper is technically sound and the results are consistent, I feel it falls slightly below the acceptance bar in terms of novelty and impact.
- 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.
After reading the rebuttal, I would change my final recommendation to Accept.
The rebuttal addressed several of my concerns. In particular, the comparison with raw volume and z-score volume conditioning makes VCS more convincing, and the Gen–Train versus Ref–Train nearest-neighbor analysis partly addresses my concern about sequential error accumulation. The clarification of the MAISI comparison and runtime estimate was also useful.
I still think the novelty is moderate, and the practical utility is mostly shown through controllability and distribution matching rather than downstream clinical or simulation workflows. However, the rebuttal makes the method better justified than I initially thought, so I now view the paper as slightly above the acceptance threshold.
Author Feedback
We thank the reviewers and address each concern below. Several details were omitted for space and are summarized here. We evaluate each organ by comparing Gen–Train and Ref–Train nearest-neighbor distances, which better assess manifold realism than pairwise Dice because plausible generations need not voxel-match held-out references. For 9/11 organs, the Gen–Train vs. Ref–Train gaps remain small, within ±1.5 mm HD95 and ±1 mm Chamfer, including late-stage organs, e.g., pancreas (+0.35/+0.29), rectum (-0.54/-0.43), duodenum (-0.66/-0.46), and bladder (-1.43/-0.92). Two exceptions are kidneys (+4.59/+2.63) and colon (+7.20/+4.64), both of which reflect representation limitations rather than failure of the sequential formulation. The kidneys are modeled as a single aggregate label despite being disconnected and often asymmetric left/right structures, likely contributing to the Dice spread of 0.60±0.23.For the colon, SDF representation is less suitable for long tubular anatomy, leading to path discontinuities, while nearest-neighbor distance is also sensitive to natural trajectory variation. We trained the same architecture with raw volume, z-score volume, and VCS, then swept the condition and measured liver volume. Raw volume conditioning performed poorly, with mean absolute error (MAE) of 460.3 mL, slope 0.39, and coefficient of determination (R²) 0.13.Z-score conditioning improved control but remained limited, with MAE 286.3 mL, slope 0.59, and R² 0.50.VCS performed best, with MAE 239.1 mL, slope 0.74, and R² 0.70, after converting to target volumes for evaluation. Lower MAE indicates more accurate volume targeting, a slope closer to 1 indicates a stronger response to the requested control value, and higher R² indicates more consistent calibration. For weakly correlated organs, the linear habitus model effectively reduces VCS to a centered/scaled volume, making it harmless while preserving a standardized control across organs. More complex nonlinear habitus models may help, but these results already show that VCS is stronger than direct volume conditioning. For MAISI, we used size = 0.9 because it produced valid, diverse high-volume livers; the comparison is not sensitive to this, since size = 1.0 collapses diversity substantially (pairwise Dice 0.957, Chamfer 3.71 mm), while size = 0.8 maintains diversity (Dice 0.802, Chamfer 13.37 mm) but remains below AbdomenGen. Very small sizes such as 0.1 or 0.2 produced partially cropped livers and are not meaningful anatomical baselines. In related work, Konz, Zhuang, Wang, and Belton primarily target CT/MRI image-volume synthesis, making them related but not directly comparable to explicit shape generation. GuideGen is closest because it produces masks and paired CT from text, but its control is semantic/textual rather than calibrated organ-wise scalar shape control. Finally, we agree that denoising backbone and FiLM conditioning are standard. The contribution is the calibrated formulation: sequential organ-wise shape synthesis with a habitus-decoupled control variable for distribution-level anatomical matching. This makes AbdomenGen shape-first, organ-wise, and calibration-focused. Although volume is a coarse descriptor, it is clinically meaningful and can be extended to other scalar shape descriptors. For error accumulation, our nearest-neighbor analysis is computed end-to-end on the full generated sequence. Thus, late-stage organs such as the pancreas, rectum, and duodenum remaining within natural variability suggest that sequential error propagation is bounded. The hepatomegaly experiment is a population-level utility test, showing that the results are not just plausible but rather can match a clinically meaningful target distribution. Computationally, a full 11-organ generation takes ~7 min/patient on an RTX A5000 with 10-step DDIM sampling (~38 s/organ). Since synthesis is sequential, peak memory remains at organ-level, supporting use in resource-constrained settings.
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.
Reviewers agree that the paper tackles a relevant problem and the framework is technically sound. However authors should address in the rebuttal gaps in validation, novelty, and how the paper positions itself to other works. This includes a better analysis of the complete multi-organ evaluation (all 11 organs), explaining failure cases (e.g., kidneys), and justifying design choices such as the sequential generation order. Furthermore, the novelty and relation to prior work must be clarified, especially to recent controllable generation methods. Finally, the authors should better justify the practical utility of the method, including robustness, error propagation in sequential generation.
- 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.
Following author’s rebuttal, there is overwhelming consensus from the reviewers supporting acceptance of this paper.
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.
This paper proposes an interesting method and provides a detailed analysis of the experimental results in the evaluation. I believe this paper is worthy of acceptance.
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.
The reviewers unanimously provided positive recommendations, and I agree with their decision.