acceptodds
Under review as a conference paper at ICLR 2027

Counterfactual 3D Pelvic Fracture CT Synthesis via Anatomy-Guided Geometric Priors and Region-Aware Latent Diffusion

Abstract

Pelvic fractures are severe traumatic orthopedic injuries that exhibit an extreme long-tail data distribution, severely limiting the generalization of deep learning segmentation models on rare but clinically critical subtypes, such as Denis type-III sacral fractures. Unlike pathological anomalies that merely manifest as local texture changes, fractures involve complex geometric discontinuities of cortical bone and rigid displacements of detached fragments. Synthesizing such injuries requires anatomically valid fracture masks, but relying solely on real, scarce fracture annotations severely limits augmentation diversity. We address this by reformulating fracture synthesis as a local counterfactual editing task: generating a fracture on a healthy pelvic CT while strictly preserving the original soft tissue background. Our framework introduces a Geometric Fracture Simulation Module (GFSM) to construct explicit clinical topology priors directly from healthy bone masks, effectively decoupling the generation process from annotation scarcity. A Mask-Conditioned Latent Diffusion Model (MC-LDM) then synthesizes high-fidelity fracture CTs based on these geometric priors, optimized by a region-aware composite loss to preserve fine structural details. To ensure seamless structural integration, we employ a novel ghost-erasure and 3D Gaussian feathering pipeline. Evaluated on PENGWIN2024 and CTPelvic1k datasets, our method achieves state-of-the-art generation quality. Crucially, augmenting downstream 3D segmentation with our synthetic data improves the Dice score on Denis type-III fractures from 0.0291 to 0.6342 and reduces HD95 from 94.7068 mm to 26.3188 mm, directly demonstrating the clinical value of geometry-grounded synthesis in the long-tail trauma regime. Code and data: https://anonymous.4open.science/r/Counterfactual-Fracture-Synthesis-6DD3/

Primary area:
generative models
Keywords:
Fracture CT Synthesis, Latent Diffusion, Counterfactual Generation, Geometric Priors, Long-Tail Augmentation

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_r8mnOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes synthesizing 3D pelvic fracture CTs by combining template-based geometric fracture simulation with mask-conditioned latent diffusion. It constructs displaced bone fragments from healthy anatomy, generates their CT appearance, and blends them into the original scan. Augmenting training with these synthetic samples substantially improves rare-fragment segmentation

Critical strengths
  1. Using generative model to augment fraction segmentation is well-motivated given the scarily of the fraction data.
  2. The use of template to generate fracture is interesting and can provide realistic-like images.
Critical weaknesses
  1. The rigid-transform operation is described, but numerical rotation/translation ranges and severity-dependent sampling distributions are not reported. Their effects on anatomical plausibility and downstream performance remain unclear.
  2. The appendices provide qualitative examples, but systematic analysis of template diversity, anatomical alignment, fragment overlap, and volume preservation is missing.
  3. Whole-volume image metrics and downstream segmentation provide useful evidence, but expert assessment and localized evaluation of fracture interfaces are needed to substantiate realism claims.
Questions for authors
  1. Could the authors specify the rotation/translation ranges and severity-dependent sampling distributions, and assess how these choices affect anatomical plausibility and downstream segmentation?
  2. Could the authors quantify template diversity and evaluate anatomical alignment, fragment overlap, and volume preservation across the generated samples?
  3. Could the authors provide expert assessment of fracture realism and fracture-region-specific evaluation?
2 comments

We thank the reviewer for the focused and precise technical questions. We address each in full.

Weakness 1 / Question 1: Rotation/translation ranges and severity-dependent sampling distributions are not reported.

We report the exact sampling ranges used by GFSM's rigid cut-and-displace step: rotation angles (α, β, γ) are drawn uniformly from ±10° per axis, and translation (t) uniformly from ±10 voxels per axis (realized ranges across all generated samples: rotation ≈ [−9.7°, 9.8°], translation ≈ [−15.0 mm, 14.1 mm]). Anatomical plausibility is enforced via rejection sampling: candidates are discarded if the displaced fragment touches the volume boundary, if volume retention falls below 98% (realized: 99.5–100.2%), or if fragment overlap exceeds 1% (all accepted samples satisfy this constraint, i.e., no inter-fragment collision was observed).

We also clarify an error in Section 3.4: the original text states that "sampling distributions for (α, β, γ, t) are constrained by injury severity," which is inaccurate. In the current implementation, these parameters are sampled from a single fixed range independent of severity; severity currently determines only which anatomical template is retrieved in Stage 2, not the transform magnitude in Stage 3. We have corrected Section 3.4 and added a limitation statement in Section 5. Extending severity-conditioned sampling to comminuted multi-fragment fractures is left for future work, contingent on annotated comminuted-fracture data.

Weakness 2 / Question 2: Template diversity and anatomical alignment are not quantified.

Regarding template diversity: by construction, GFSM's Fracture Surface Atlas (Stage 1) mines exactly one surface template per real training case; template diversity per category is therefore bounded by the real training-sample counts already reported (Pubic: 33, Denis I: 16, Denis II: 11, Acetabular: 5, Iliac: 5, Denis III: 2). We now state explicitly that every Denis III synthetic sample is generated from only 2 distinct real templates — a direct consequence of the data scarcity that motivates this work, and we list this as a limitation in Section 5.

Regarding anatomical alignment: Stage 2 retrieval selects templates by exact match on anatomical site and fracture type, so the zone label of every generated fracture is inherited from a real case already confirmed to be in that zone. Stage 3's rigid transform displaces the fragment locally (realized translation ≤15 mm), which is small relative to zone extents and insufficient to relocate a fracture into a different anatomical region. We consider zone correctness preserved by construction, supported by the volume retention (99.5–100.2%) and zero inter-fragment collision results above.

Weakness 3 / Question 3: Expert assessment and localized evaluation of fracture interfaces are needed.

We do not currently have access to a clinical collaborator to conduct a blinded expert evaluation within the rebuttal window. As partial substitutes, we provide: (1) the quantitative geometric validity checks reported above (volume retention, fragment overlap); (2) the staged pipeline ablation demonstrating that each compositing step contributes measurably to downstream task performance; and (3) the existing qualitative multi-planar visualizations in Appendix D/F. We agree that a clinician-rated realism study would strengthen the realism claim and will pursue this with a radiology collaborator as follow-up work, which we have added to the limitations section.


Dear Reviewer r8mn,

As the discussion deadline is approaching, we would like to kindly follow up. We have responded to all of your questions, including the exact rotation/translation ranges, volume preservation and fragment-overlap statistics, template diversity, and anatomical alignment. We would appreciate your feedback on whether our response addresses your concerns, and we would be happy to answer any further questions.

Thank you for your time and constructive review.

Reviewer_SjyGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method for augmenting/generating CT-Segmentation fracture pairs using combination of traditional geometric+inpainting followed by task specific diffusion-based generative model for plausible looking fractured CT texture. The plausibility of such generation is measured using automated semantic image metrics such as LPIPS, FID, SSIM and PSNR compared against other generative appraoches such as GAN, LDM. Improvement in downstream CT Pelvic Bone segmentation task is also shown via synthetic fracture generation used for augmenting the training set.

Critical strengths
  • The problem is well motivated and a plausible fracture generation mechanism is proposed that combines geometric and diffusion-based generative modeling.
Critical weaknesses
  • I think the ablation is limited. The paper only considers VAE reconstruction losses for representation learning in the ablation study. Various Components of the fracture synthesis and design decisions could use ablation. Additionally, I would suggest a qualitative Human evaluation by Orthopedic Expert to support the generation fidelity argument better.
  • Across the paper, the presentation is somewhat oversold. The 21× improvement on Denis type-III small fractures is correct, but only a few training samples are available for training and for generating additional fractures, so the sample size is too small to consider the improvement significant; the number of type-III test cases is also unclear. "Counterfactual" is used to mean conditional fracture synthesis, and the "text-guided" retrieval and "clinical text descriptors" are barely described and seem to be fixed site/fracture-type labels inferred from segmentation masks, i.e. a category lookup.
  • The downstream segmentation evaluation is weak. A 3D U-Net baseline is acceptable, but the PENGWIN challenge documents stronger fracture-aware methods, such as FracSegNet's distance-weighted loss would be a more meaningful reference. The paper compares against none, so the gain from synthetic data is not placed against the existing state of the art. The synthesis method need not be state of the art, but comparison with the literature and the leaderboard is missing and feasible, since public implementations exist.
Questions for authors
  • Beyond the VAE reconstruction-loss ablation, could you ablate the other pipeline components (the region-aware loss, ghost erasure and z-axis smoothing) to quantify their contribution to generation quality and/or to the downstream gain?
  • On which dataset and split is the Table 3 segmentation evaluated (PENGWIN, CTPelvic1K or both; split per patient or per sub-unit)? Please add a column with the number of evaluation samples for each subtype, and confidence intervals across seeds.
  • PENGWIN and CTPelvic1K do not seem to provide fracture category labels. How were the categories obtained (automated or manual, by whom, with what agreement), and how many Denis type-III cases are in the training and test sets?
  • Could you please try better baseline fracture-aware segmentation methods from the PENGWIN challenge, such as FracSegNet's distance-weighted loss to test the complementary nature of the realistic synthesis and put this in context with the leaderboard methods.
  • What does "anatomically matched healthy samples" mean in the paper? On what basis are the 153 Samples matched?
  • The datasets do not seem to additional clinical text. Are "text-guided" retrieval and "clinical text descriptors" just fixed site/fracture-type labels inferred from segmentation masks, i.e., a category lookup?
  • The anonymous repository is empty. Will the code, data splits and seeds be released during the discussion period?
2 comments

We thank the reviewer and address each point below.

Weakness 1: Ablation is limited; only VAE reconstruction losses are ablated. Other pipeline components (region-aware loss, ghost erasure, z-axis smoothing) should be ablated.

We ran a full staged ablation during rebuttal, covering ghost-erasure, MC-LDM texture, and z-axis feathering, beyond Table 4's VAE-GAN loss ablation:

MethodDenis III SmallSacrum avgAll-6 avg
Baseline (no augmentation)0.0290.1280.347
Naive geometric augmentation0.1120.2110.359
Ablation A: no ghost-erasure0.0000.0010.094
Ablation C: GFSM geometry only, no MC-LDM texture0.4890.3280.475
Ablation B: no z-axis feathering0.5600.5170.479
Full method (GFSM + MC-LDM)0.6340.7340.682
Most informative: removing ghost-erasure (A) pulls performance below Baseline (0.094 vs. 0.347 all-6-avg) — ghost-erasure is a prerequisite, not cosmetic; without it, residual cortical halos actively harm downstream segmentation rather than merely failing to help. GFSM alone (C) already recovers 0.489, confirming the geometric prior drives most of the benefit, with MC-LDM contributing the remainder.
Weakness 2: Presentation is somewhat oversold (21× improvement with very small n; sample sizes unclear).

We agree and address this directly:

CategoryTrain (before)Train (after)Test (real)
Pubic3333 (1.0×)9
Iliac530 (6.0×)2
Acetabular530 (6.0×)2
Denis I1630 (1.9×)3
Denis II1130 (2.7×)3
Denis III230 (15.0×)1

We also report paired Wilcoxon p-values across all 20 test cases (p=0.024 main, p=0.015 small). The 21× claim for Denis III is accurate but based on n=1 test case; we do not claim significance for individual rare subtypes.

Weakness 3 / Question: "Counterfactual" and "text-guided" seem oversold.

We agree and are revising both terms. "Counterfactual" follows prior work we cite (Wolleb et al. 2022; Fontanella et al. 2025): a conditional generation asking "what would this healthy scan look like under a given injury," not a causal intervention in Pearl's sense; we will clarify this in Section 3.4. "Text-guided": input T is a structured clinical descriptor (site + fracture type + severity), not free-form text — retrieval is exact-label matching, not semantic parsing. We will rename this "attribute-guided retrieval" and list free-text parsing as future work.

Question: On which dataset and split is Table 3 evaluated?

All results in Tables 1–3 are evaluated entirely within PENGWIN2024; CTPelvic1k only supplies healthy background CTs for counterfactual generation (Section 4.1), not segmentation evaluation. The train/val/test split (7:2:1) is at the sub-unit level (sacrum / left-hip / right-hip), matching PENGWIN2024's own partition of each scan into three independently evaluated sub-units.

Question: How were fracture category labels obtained?

The 153 real fracture cases were manually classified into AO/OTA categories (hip: pubic/iliac/acetabular) and Denis categories (sacrum: Zone I/II/III) from the fragment segmentation masks, per official clinical criteria (Meinberg et al., 2018; Denis et al., 1988). One author labeled all cases; a second team member independently reviewed ambiguous cases. We will add this detail to Section 4.1.

Question: Try FracSegNet's distance-weighted loss as a stronger baseline.

We evaluated nnU-Net (self-configuring, state-of-the-art) on both Baseline and Augmented data, reporting all effects including weak/negative ones: Baseline already performs strongly on scarce data (Denis III small-fragment Dice 0.836 vs. 3D U-Net's 0.029), showing a well-engineered recipe can partly compensate for scarcity. Averaged across categories, augmentation still yields a modest net gain (0.785→0.827, +0.042): large gains on Pubic (+0.099), Denis I (+0.104), Acetabular (+0.291), a decrease on Iliac (−0.220 — only 5 real training cases pre-augmentation, our highest-variance category, so we cannot rule out variance over a systematic effect), small decreases on Denis II/III. We do not claim backbone-invariant benefit; comparison with fracture-aware losses like FracSegNet remains future work.

Question: What does "anatomically matched" mean for the 153 healthy samples?

The 153 healthy samples from CTPelvic1k mirror the regional composition of our 153 real fracture samples: 43 sacrum, 55 left-hip, 55 right-hip sub-units. We did not match at the patient level (age, sex, bone morphology); "anatomically matched" refers to this region-count correspondence only. We will revise Section 4.1 to state this explicitly.

Question: Anonymous repository is empty. Will code, data splits, and seeds be released?

We apologize for the delay. Code, data splits, and seeds are being uploaded during the discussion period and will be complete before the deadline.


Thank you to the author for the useful ablation study, experiments and detailed rebuttal.

Regarding the ablation, the geometry-only module although seem to "drive most of the benefit" for Denis III (n=1), the claim may need to be toned down for the other cases.

Few followup to the earlier questions and rebuttals:

  • Question: How were fracture category labels obtained?

The 153 real fracture cases were manually classified into AO/OTA categories (hip: pubic/iliac/acetabular) and Denis categories (sacrum: Zone I/II/III) from the fragment segmentation masks, per official clinical criteria (Meinberg et al., 2018; Denis et al., 1988). One author labeled all cases; a second team member independently reviewed ambiguous cases. We will add this detail to Section 4.1.

The category labels being generated and reviewed by the authors may somewhat make sense for basic anatomy labelling but denis zoning seem to require quite a clinical background and understanding relationship between structures, the accuracy of which may affect template retrieval. Please state the labeller's clinical qualification if any, and if not, the agreement between readers for basic sanity check. Please consider providing clinician-validated labels in future work.

  • Question: On which dataset and split is Table 3 evaluated?

All results in Tables 1–3 are evaluated entirely within PENGWIN2024; CTPelvic1k only supplies healthy background CTs for counterfactual generation (Section 4.1), not segmentation evaluation. The train/val/test split (7:2:1) is at the sub-unit level (sacrum / left-hip / right-hip), matching PENGWIN2024's own partition of each scan into three independently evaluated sub-units.

The splitting of the dataset at "sub-unit level (sacrum / left-hip / right-hip)" does not mimic a clinical workflow. The original PENGWIN challenge splits at the patient-level although they may report sub-unit level metrics in addition to the whole CT scan metrics. Please report a patient-level split, number of distinct CTs per split and whether any CT appears in more than one split.

  • Question: Try FracSegNet's distance-weighted loss as a stronger baseline.

We evaluated nnU-Net (self-configuring, state-of-the-art) on both Baseline and Augmented data, reporting all effects including weak/negative ones: Baseline already performs strongly on scarce data (Denis III small-fragment Dice 0.836 vs. 3D U-Net's 0.029), showing a well-engineered recipe can partly compensate for scarcity. Averaged across categories, augmentation still yields a modest net gain (0.785→0.827, +0.042): large gains on Pubic (+0.099), Denis I (+0.104), Acetabular (+0.291), a decrease on Iliac (−0.220 — only 5 real training cases pre-augmentation, our highest-variance category, so we cannot rule out variance over a systematic effect), small decreases on Denis II/III. We do not claim backbone-invariant benefit; comparison with fracture-aware losses like FracSegNet remains future work.

The nnUnet results are informative and weaken the 21x improvement claim considerably since it corresponds to the weak 3D U-Net baseline.

Reviewer_V7NsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a framework for synthesizing 3D pelvic fracture CT images to address the shortage of training data for rare fracture types. The method uses anatomy-guided geometric simulation to generate fracture masks from healthy bone structures, followed by a mask-conditioned latent diffusion model to synthesize fracture images while preserving surrounding tissues. A region-aware reconstruction loss and an image fusion procedure are introduced to improve fracture boundaries and reduce artifacts. Experiments on PENGWIN2024 and CTPelvic1k show that the generated images achieve competitive synthesis quality and improve downstream fracture segmentation, particularly for underrepresented fracture types such as Denis type-III sacral fractures.

Critical strengths

The paper addresses an important challenge in medical image synthesis by targeting rare pelvic fracture types with limited training data. The use of anatomy-guided geometric priors provides a practical way to generate fracture configurations while preserving surrounding anatomical structures. The experiments demonstrate substantial improvements in downstream segmentation for underrepresented fracture categories, suggesting that the synthetic data can be useful for addressing class imbalance in medical imaging.

Critical weaknesses

The main weakness of this paper is its limited methodological novelty, as the proposed framework largely combines existing geometric simulation, latent diffusion, and image fusion techniques. The experimental evaluation lacks comparisons with simpler augmentation strategies and sufficient ablation studies to establish the contribution of each component. Although the downstream segmentation improvements are substantial, the extremely small number of samples in rare fracture categories raises concerns about the reliability and generalizability of these results. Furthermore, the improvements in image synthesis quality over existing diffusion-based methods are relatively modest, making it difficult to justify the additional complexity of the proposed framework.

Questions for authors

The authors should clarify whether the downstream segmentation improvements result from the proposed synthesis framework or simply from increasing the number of training samples. Comparisons with conventional geometric augmentation are needed to establish the necessity of diffusion-based synthesis. Additional ablation studies should isolate the contributions of GFSM and MC-LDM to downstream performance. Given the extremely limited number of real samples for rare fracture types, stronger evidence of experimental stability is also needed to support the reported improvements.

2 comments

We thank the reviewer for the thorough analysis. The reviewer correctly identifies limited novelty as the primary concern; our contribution lies in three specific design decisions: (1) GFSM, which decouples fracture-topology generation from scarce annotated data via attribute-guided template retrieval and rigid cut-and-displace simulation — no prior work addresses this bottleneck this way; (2) the region-aware composite loss (ROI-weighted adversarial loss + 3D gradient loss) targeting fracture-surface sharpness; (3) the Disguise-Erase-Restore compositing pipeline, which the ablation below shows is load-bearing, not cosmetic. Section 1 has been revised accordingly.

Weakness 1 / Question: Whether improvements come from the synthesis pipeline or simply from increasing the number of training samples; comparisons with conventional geometric augmentation are needed.

We ran a controlled experiment to test this directly: a naive augmentation baseline using standard random affine/elastic deformation, matched to the same per-category sample target (up to 30), training an otherwise identical 3D U-Net. Denis III small-fragment Dice (headline rare-category metric): Baseline 0.029 → Naive augmentation 0.112 → Full method 0.634. Naive augmentation recovers only a small fraction of the gap, confirming the gain is not attributable to sample count alone. Naive augmentation also degrades sacral main-fragment Dice (0.749→0.215 avg), consistent with distortion-induced confusion from unconstrained warps — a failure mode our geometry-constrained GFSM avoids by design.

Weakness 2 / Question: Ablation studies should isolate the contributions of GFSM and MC-LDM.

We ran a full staged ablation isolating each component:

MethodDenis III SmallSacrum avgAll-6 avg
Baseline (no augmentation)0.0290.1280.347
Naive geometric augmentation0.1120.2110.359
Ablation A: no ghost-erasure0.0000.0010.094
Ablation C: GFSM geometry only, no MC-LDM texture0.4890.3280.475
Ablation B: no z-axis feathering0.5600.5170.479
Full method (GFSM + MC-LDM)0.6340.7340.682

Most striking: removing ghost-erasure (Ablation A) pulls performance below the no-augmentation Baseline (0.094 vs. 0.347 all-6-class avg; 0.001 vs. 0.128 sacrum avg), showing ghost-erasure is a prerequisite, not an optional refinement — without it, residual cortical halos from healthy bone confound the conditioned generation and actively harm downstream segmentation. GFSM geometry alone (Ablation C) already recovers 0.489, confirming the geometric prior accounts for most of the gain, with MC-LDM's learned texture and compositing refinements contributing the remainder.

Weakness 3: Small sample sizes for rare categories raise reliability concerns.

We fully acknowledge this — it is the paper's central motivation. Paired Wilcoxon signed-rank tests on all 20 test cases: main-fragment Dice 0.838→0.906 (p=0.024), small-fragment Dice 0.461→0.697 (p=0.015), both significant. As an internal consistency check, Pubic (n=9, zero synthetic samples added) shows no significant small-fragment change (p=1.00), supporting that gains elsewhere come from the synthetic data, not training noise. For rare subtypes (n=1–3 test cases), formal significance testing is not meaningful; we report raw per-case Main/Small Dice (Before→After) instead of hiding behind an average: Denis I — sacrum_020 0.958→0.975 / 0.134→0.847; sacrum_004 0.969→0.939 / 0.000→0.925; sacrum_028 0.600→0.945 / 0.503→0.598. Denis II — sacrum_014 0.632→0.884 / 0.000→0.678; sacrum_019 0.579→0.964 / 0.014→0.905; sacrum_015 0.888→0.821 / 0.410→0.751. Denis III — sacrum_084 0.704→0.666 / 0.029→0.634. Iliac — left_hip_043 0.948→0.987 / 0.846→0.974; right_hip_018 0.718→0.769 / 0.115→0.071. Acetabular — left_hip_092 0.954→0.951 / 0.682→0.672; left_hip_096 0.927→0.895 / 0.403→0.688. The effect is not uniformly positive (right_hip_018 small-fragment Dice decreases), which we report rather than omit.

Weakness 4: Modest improvement in image synthesis quality over existing diffusion methods.

We agree the rendering-metric gains over LDM are modest (Table 1) — by design: our contribution is a geometry-grounded fracture-topology prior enabling downstream augmentation, not a better general-purpose CT denoiser. The complexity we introduce targets Table 3's downstream benefit, not raw rendering fidelity.


Dear Reviewer V7Ns,

As the discussion deadline is approaching, we would like to kindly follow up. We have responded to all of your concerns, including new experiments comparing against standard geometric augmentation, a staged ablation isolating each pipeline component, and paired statistical tests with per-case results for the rare categories. We would greatly appreciate your feedback on whether these address your concerns, and we would be happy to discuss any remaining questions.

Thank you for your time and thoughtful review.

Reviewer_C8pVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper propose a new method for bone fracture CT generation with LDM. It did not propose new module, but instead combined a few existing modules and achieve a significant Dice improvement. (If any module is novel, the authors should clearly say it in Contribution subsection.)

Critical strengths

The Dice improvement is large compared with comparison methods

Critical weaknesses
  1. It did not propose new module, but instead combined a few existing modules. If any module is novel, the authors should clearly say it in Contribution subsection.
  2. The comparison method is old, 2024. It is hard to believe that there is no more advanced CT diffusion model released since then. This makes the big improvement looks weak.
Questions for authors

N/A

2 comments

We appreciate the reviewer's recognition of the large Dice improvement and the constructive critique on novelty framing and baseline recency.

Weakness 1: No new module proposed; existing modules combined without clear novelty statement.

We respectfully agree that individual building blocks (latent diffusion, VAE-GAN, inpainting) are established techniques, and we acknowledge that our original contribution statement was insufficiently precise. We have revised Section 1 to explicitly state three specific contributions: (1) GFSM, which constructs fracture topology masks via attribute-guided retrieval of real fracture-surface templates combined with rigid cut-and-displace simulation — this decouples topology generation from scarce fracture annotations in a way no prior work addresses; (2) the region-aware composite loss (ROI-weighted adversarial loss + 3D gradient loss) tailored to fracture-surface sharpness; and (3) the Disguise-Erase-Restore pipeline for seamless, background-preserving 3D compositing. We have been careful not to claim the overall pipeline architecture as the novelty, but rather these three specific design decisions within it.

Weakness 2: Comparison methods are from 2024; more recent CT diffusion models may exist.

We thank the reviewer for this important point. We attempted to integrate and train a comparison against 3D MedDiffusion (2025), a more recent diffusion backbone, under our mask-conditioned setup. Due to the limited rebuttal window and compute constraints, we were unable to complete this training in time. We note that our existing comparison already includes LDM as a strong, widely-adopted baseline, against which our rendering-metric performance is comparable (Table 1); our primary contribution is the geometry-grounded GFSM prior and its downstream augmentation benefit (Table 3), not outperforming every possible diffusion backbone. We commit to adding a comparison with more recent generative backbones in the camera-ready version.


Dear Reviewer C8pV, As the discussion deadline is approaching, we would like to kindly follow up. We have responded to both of your concerns: we revised the contribution statement to specify what is novel in our work, and we explained the status of the comparison with more recent diffusion backbones. We would be grateful to hear whether our response addresses your concerns, and we would be happy to answer any further questions you may have. Thank you for your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.