List of Papers Browse by Subject Areas Author List
Abstract
Unsupervised anomaly detection (UAD) offers a promising direction for fracture analysis where abnormal annotations are scarce, yet pelvic X-ray fracture detection remains exceptionally challenging due to overlapping anatomy, projection artifacts, and the inherently asymmetric presentation of fracture patterns. We introduce SymmDiff, a symmetry-aware diffusion framework that embeds contralateral anatomical consistency directly into the generative correction process, mirroring the clinical practice of bilateral comparison. By pairing anatomically corresponding left–right regions and enforcing symmetry-consistent reconstruction during diffusion, SymmDiff generates a normal-corrected radiograph that selectively suppresses fracture-induced asymmetries while preserving global anatomical structure, enabling precise fracture localization through residual discrepancies between the input image and its symmetry-corrected counterpart. Extensive evaluations on pelvic radiographs spanning subtle to severe fractures show that SymmDiff consistently outperforms existing anomaly-detection baselines, highlighting the effectiveness of symmetry-guided generative modeling for reliable fracture localization in clinical imaging. Code is available at https://itsabdulrahman.github.io/symmdiff.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4114_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://itsabdulrahman.github.io/symmdiff/
Link to the Dataset(s)
N/A
BibTex
@InProceedings{RahAbd_SymmDiff_MICCAI2026,
author = { Rahman, Abdul AND Ghafoor, Afnan AND Lee, Bumshik},
title = { { SymmDiff: Symmetry-Guided Diffusion for Fracture Localization in Pelvic Radiographs } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
In this paper, the authors introduce a novel symmetry-guided diffusion model for fracture localization in pelvic radiographs. Specifically, the method embeds contralateral anatomical information into the model to detect fractures.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The proposed method is novel. The idea of using symmetry-corrected generated images to identify fracture locations is interesting.
- 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 proposed method achieves good performance compared with other state-of-the-art generative models, it would be better if the authors also compared it with commonly used detection methods, such as YOLO. Fracture detection may not necessarily require generative models like DDPM for this task. In addition, the paper lacks evaluation metrics for image quality, such as the FID score. It would also be beneficial to include radiologists to assess the quality of the generated images.
- 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 paper presents a novel and interesting approach by introducing a symmetry-guided diffusion model for fracture localization in pelvic radiographs. The idea of leveraging contralateral anatomical symmetry and using symmetry-corrected generated images to highlight fracture regions is both intuitive and technically interesting. This formulation provides a new perspective on incorporating anatomical priors into generative modeling for medical imaging tasks, which could inspire further research in this direction.
The experimental results demonstrate competitive performance compared to other state-of-the-art generative models, suggesting that the proposed approach is effective for the task. Despite some limitations—such as the lack of comparison with standard detection methods (e.g., YOLO), absence of image quality metrics like FID, and missing human expert evaluation—the core idea is novel and shows clear potential impact.
Overall, while the experimental validation could be strengthened, the methodological contribution and innovation justify a weak accept, as the work is likely to stimulate further research and improvements in symmetry-aware medical image analysis.
- 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 #2
- Please describe the contribution of the paper
The paper introduces a generative framework for unsupervised anomaly detection (UAD) in pelvic X-rays. It addresses the difficulty of identifying fractures amidst complex overlapping anatomy and projection artifacts by mimicking the clinical practice of side-to-side comparison. The model uses a symmetry-conditioned latent diffusion process that establishes a patient-specific representation of normal anatomy through contralateral correspondence.
- 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 method is clinically relevent and the Symmetry-Diffusion appoarch is innovative. By embedding contralateral consistency directly into the denoising process, the model suppresses asymmetries caused by normal anatomical variation or projection tilt.
- 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.While the pelvis is generally symmetric, true biological asymmetry and variations in patient positioning (tilt/rotation) during X-ray acquisition are common. Although the authors suggest that latent encoding provides robustness to ‘minor projection mismatch’, the paper would benefit from a more detailed analysis or a ‘stress test’ showing how the model performs under significant pelvic tilt or in patients with pre-existing non-fracture asymmetries.
2.A primary concern for any symmetry-guided model is the occurrence of bilateral fractures. If both sides of the pelvis are fractured, the ‘structural prior’ from the contralateral side may itself be abnormal. While the authors mention using identity constraints and population-level atlases as future work, the current manuscript lacks a clear demonstration of how the model fails or adapts when symmetry is violated on both sides simultaneously.
3.The qualitative results show some fractures that appear relatively obvious visually. To better justify the necessity of this complex generative framework, the authors may provide a more focused evaluation on ‘subtle’ or ‘occult’ fractures where standard radiological interpretation often fails. Quantifying performance specifically on these ‘difficult’ cases would more clearly highlight the clinical value of the SymmDiff approach.
- 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 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 framework relies heavily on an assumption of anatomical symmetry, which may fail in cases of bilateral fractures or significant physiological asymmetry. Furthermore, while the method performs well on “normal-to-normal” completion, its effectiveness in correcting actual severe clinical deformities remains unproven. The reliance on a “black box” transformer also lacks the interpretability of traditional statistical models, and the current evaluation lacks a comparison against a version of itself guided by a general healthy atlas rather than patient-specific symmetry, making it difficult to fully isolate the unique value of the symmetry-inductive bias.
- 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 #3
- Please describe the contribution of the paper
This paper introduces SymmDiff, a symmetry-guided latent diffusion framework for unsupervised anomaly detection (UAD) in pelvic radiographs. The core idea is to embed contralateral anatomical consistency directly into the reverse diffusion process, mirroring the clinical practice of bilateral comparison. The input image and its horizontally flipped counterpart are encoded into a shared latent space via a fine-tuned VAE. During inference, an ROI mask derived from bilateral discrepancy identifies suspicious regions, a latent swap initializes those regions toward the contralateral reference, and symmetry-conditioned DDIM denoising reconstructs a pseudo-healthy image with reverse updates constrained to the ROI. Anomalies are localized through a multi-scale fusion of pixel, gradient, perceptual, and latent residuals between the input and reconstruction. Evaluation on a single-center pelvic X-ray dataset (1,268 normal training cases, 534 abnormal test cases) shows consistent improvements over four diffusion-based UAD baselines at both image and pixel levels, and ablations demonstrate that each component contributes to performance.
The overall method appears innovative; however, the authors do not describe in sufficient detail several parts of the proposed architecture that are essential for reproducing the results, such as the type of VAE encoder used and the architecture of the U-Net network.
- 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 central methodological idea — conditioning each reverse diffusion step on a patient-specific contralateral reference — is novel within the diffusion-based UAD literature. Prior methods either treat denoising as anatomy-agnostic or apply bilateral comparison post-hoc or in discriminative settings. Embedding symmetry into the reverse trajectory dynamics is a genuinely new contribution that is well-motivated by established clinical reading practice for pelvic radiographs.
The clinical rationale is strong. The pelvis is one of the clearest anatomical cases for bilateral comparison, and the paper connects this practice to the technical design in a coherent way. The choice of application domain is not arbitrary — it is one where the proposed inductive bias is most naturally justified.
The ablation study is thorough and informative. Five ablations each isolate a specific component, and all show consistent degradation. Importantly, the removal of symmetry conditioning causes the largest performance drop, which confirms that the core mechanism is driving the results rather than ancillary design choices. This is the most convincing evidence in the paper.
The ROI-constrained reverse diffusion (Eq. 9) is cleanly formulated with a verifiable identity preservation property outside the ROI. This targeted correction strategy is a sensible design choice that reduces the risk of global structural drift during reconstruction, addressing a known failure mode of diffusion-based UAD.
The performance gains over the strongest baseline (THOR) are substantial, particularly at the pixel level (AUPRC 36.2 –> 43.9, F1max 37.2 –> 42.7), suggesting the method provides meaningfully better localization rather than marginal improvement.
- 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 comparison set is narrow and limited exclusively to diffusion-based UAD methods. There is no comparison with feature-based anomaly detection approaches such as PatchCore, which is cited (ref [4]) but not benchmarked. More critically, there is no comparison with a simpler symmetry baseline — for instance, left-right subtraction with alignment and thresholding. Without this, it is impossible to disentangle how much of the performance gain comes from the diffusion framework versus the symmetry prior itself. This is a significant omission that weakens the empirical contribution.
The evaluation is conducted on a single anonymized dataset from one institution with no external validation. Pelvic radiographs vary considerably across institutions due to equipment, positioning protocols, and patient demographics. The symmetry assumption can degrade under oblique projections, rotated patients, or bilateral pathology. The authors acknowledge cross-site generalization as future work, but the absence of any robustness analysis on these factors limits confidence in the method’s generalizability. There is no stratification of results by fracture severity, location, or type. With 534 abnormal cases, meaningful subgroup analysis is feasible. This matters because the clinical value proposition differs substantially for subtle versus severe fractures — if the aggregate numbers are driven primarily by severe cases that radiologists would not miss, the practical utility is limited. Five qualitative examples do not substitute for systematic analysis.
The latent swap claim (Eq. 8) rests on an unstated assumption that the contralateral latent is closer to the true normal latent than the original input within ROI regions. This is plausible for unilateral fractures but could fail for bilateral pathology, symmetric normal variants, or significant projection asymmetry. The paper does not acknowledge or bound this assumption.
No confidence intervals, standard deviations, or statistical significance tests are reported for results. Given the dataset size, some estimate of variance would strengthen the claims considerably.
The authors do not describe in sufficient detail several architectural components that are critical for reproducibility. Specifically, the type of VAE encoder and the architecture of the U-Net network are not specified, making it challenging to independently reimplement the method even with access to the described training hyperparameters.
- 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
The most impactful improvement would be the inclusion of non-diffusion baselines. A feature-based method such as PatchCore and a simple symmetry-difference baseline would substantially strengthen the evaluation by clarifying what the diffusion framework contributes beyond the symmetry prior. This is not a request for additional experiments, but rather an observation about what would make the empirical story complete.
A stratified analysis by fracture severity would significantly strengthen the clinical relevance argument. Even a coarse split into minor, moderate, and severe categories with per-group metrics would help readers assess where the method adds value.
The bilateral pathology case deserves more than a passing mention in the conclusion. Since the entire framework rests on the assumption that the contralateral side provides a reliable normal reference, the failure mode when both sides are abnormal is fundamental. Some analysis or at least explicit discussion of how the method behaves in this scenario would be valuable.
- 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 paper presents a well-motivated and technically coherent idea — embedding pelvic symmetry directly into the reverse diffusion process — and the ablation study provides credible evidence that this mechanism is the primary driver of performance. The gains over the strongest diffusion-based baseline are non-trivial, and the clinical framing is appropriate.
However, several issues, collectively, prevent a higher score. The empirical evaluation is narrow: the absence of non-diffusion baselines (e.g., PatchCore) and, more critically, the absence of a simple symmetry-difference baseline make it difficult to isolate the contribution of the diffusion framework from that of the symmetry prior alone. The single-center evaluation without external validation or robustness analysis limits confidence in generalisability. The lack of subgroup analysis by fracture severity leaves the clinical utility of the method unclear. No statistical measures of uncertainty are reported for the results given.
Reproducibility appears to be an issue. The VAE encoder type and U-Net architecture — both fundamental to the proposed framework — are not specified. The ROI scoring function, binarization threshold, fusion weights, and patch masking details are either vague or absent. This is a key shortcoming for a methods contribution, and it is the primary factor limiting the score. Much of this may be addressed by the authors’ planned code release, but is difficult to understand if this is true from the paper alone.
- 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 thank the reviewers and AC for their insightful feedback and for recognizing the novelty of embedding contralateral consistency directly into the diffusion trajectory. Below we address the major concerns raised by the reviewers. Choice of baselines (R1, R3): SymmDiff is designed for unsupervised anomaly Detection (UAD), motivated by the scarcity of fine-grained fracture annotations and the difficulty of predefining fracture morphologies. Supervised detectors such as YOLO require annotated abnormal cases for training and therefore address a different setting from the proposed UAD formulation. We agree that PatchCore [4] is a strong feature-memory method. Our current evaluation focused on reconstruction-based diffusion methods to isolate the contribution of symmetry-guided correction within a comparable paradigm, as PatchCore takes a fundamentally different approach, scoring anomalies against a memory bank of normal patches rather than via reconstruction discrepancy. A thorough cross-paradigm comparison would meaningfully complement our evaluation and is an important direction we plan to pursue. Contribution of symmetry vs. diffusion (R3): We agree that simple left-right difference is an important conceptual baseline. The submitted ablations help isolate this point. The ROI is derived from bilateral discrepancy, but removing ROI-constrained denoising reduces performance (pixel AUROC 83.2 vs. 88.4), indicating that raw symmetry discrepancy alone is insufficient. Conversely, removing symmetry conditioning degrades the model toward an anatomy-agnostic diffusion baseline (79.9 vs. 76.5), showing that the improvement comes from embedding the contralateral prior into the reverse trajectory, not merely from using diffusion. Direct subtraction compares sides; SymmDiff uses the contralateral side to synthesize a pseudo-healthy correction while preserving non-ROI anatomy. Robustness to symmetry violations (R2, R3): Two failure modes were raised: projection tilt and bilateral fractures. The method does not assume perfect pixel-wise symmetry. The use of latent space rather than pixel space abstracts local textures into higher level features, which improves robustness to minor projection tilt and rotation. For bilateral fractures, the ROI gating with identity constraints (Eq. 9) provides a stabilizing bias: regions outside the ROI mask remain unchanged, so when both sides are abnormal the global anatomy is preserved rather than corrupted by a degraded contralateral prior. Population level healthy atlases could serve as a secondary anchor for bilateral fracture cases, which is a key area for future work. Reproducibility (R3): The VAE is the pre-trained perceptual-compression autoencoder from [12], trained with a perceptual loss, patch-based adversarial objective, and KL regularization, with a down sampling factor of 8, we fine-tune it on pelvic radiographs. The denoiser is an attention U-Net with timestep embedding and a squared-cosine-beta noise schedule, with z_sym incorporated via channel-wise concatenation as described in Sec. 3.2.Configuration files and training scripts will be released with the code.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
All three reviewers recommend weak accept. The core idea of embedding contralateral symmetry directly into the reverse diffusion trajectory is clinically well-motivated and technically novel within the diffusion-based UAD literature, and the ablation credibly demonstrates that symmetry conditioning is the primary performance driver. The authors should address in the final version the absence of a simple left-right subtraction baseline and the missing architecture details, which are critical for reproducibility.