SymmDiff: Symmetry-Guided Diffusion for Fracture Localization in Pelvic Radiographs

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.

Key Contributions

A symmetry-conditioned latent diffusion model that establishes a patient-specific representation of normal pelvic anatomy through contralateral correspondence, providing a structural prior for fracture localization.
A symmetry-conditioned reverse diffusion formulation in which contralateral information guides the reverse transition at each step, stabilizing the anatomical trajectory during iterative denoising.
An ROI-guided reverse diffusion scheme with contralateral latent-swap initialization that softly restricts updates to regions exhibiting elevated bilateral discrepancy, reducing projection-induced over-correction while preserving global anatomy.

Method

(a) Training with input x and its flipped counterpart x_sym; (b) random patch-mask generation; (c) inference pipeline with ROI latent swap and DDIM sampling; (d) left/right ROI scoring; (e) multi-scale anomaly fusion.

Symmetry-Aware Latent Encoding

The input radiograph x and its horizontally flipped reference x_sym = T(x) are encoded by a pelvic-finetuned VAE into latents z and z_sym. Because pelvic anatomy is bilaterally correspondent, the two latents are not independent, so conditioning on the contralateral side strictly reduces uncertainty (H(Z | R) < H(Z)) and supplies a patient-specific structural anchor rather than a population-level notion of normality.

Symmetry-Constrained Training

Random patch masking. Patches of the noisy latent are zeroed before denoising, forcing the U-Net to reconstruct from global structure instead of memorizing local texture.
Reference conditioning. z_sym is concatenated channel-wise with the masked latent, so the denoiser ε_θ inpaints masked regions using bilateral context. The contralateral reference therefore shapes the whole reverse trajectory, not a post-hoc residual.

ROI-Targeted Inference

Soft ROI gate. A side-to-side discrepancy score yields a mask m_ROI that narrows the correction search space without acting as a hard pre-filter.
Latent swap initialization. Inside the ROI the trajectory is initialized from the patient's own contralateral latent, biasing correction toward a symmetry-consistent manifold.
Identity outside the mask. Deterministic DDIM (η = 0) updates only inside the ROI, so the reverse process is the identity elsewhere — imperfect mask estimation cannot induce global structural drift.

Multi-Scale Anomaly Fusion

A single residual is unreliable in pelvic radiographs, where abnormalities differ in scale and cue. Four complementary discrepancies between the input and its reconstruction — pixel, gradient, perceptual (ResNet features), and latent — are normalized per map and combined convexly into one anomaly map. Thin cortical breaks surface most strongly in the gradient and perceptual views, while broader deformities dominate the pixel and latent views.

Results

94.2

Image-level AUROC

88.4

Pixel-level AUROC

+7.7

Pixel AUPRC over the
second-best method

0

Abnormal labels used
during training

Quantitative Comparison

We evaluate on a retrospective Chosun University Hospital pelvic X-ray dataset. SymmDiff is best at both image and pixel level. Relative to the second-best method (THOR), image-level AUROC / AUPRC / F1_max improve by +4.9 / +2.5 / +2.8, and pixel-level by +3.9 / +7.7 / +5.5.

Method	Image-level			Pixel-level
Method	AUROC	AUPRC	F1_max	AUROC	AUPRC	F1_max
DDPM	84.5	83.0	80.1	76.5	23.3	24.6
AnoDDPM	87.5	85.4	83.0	79.5	26.1	27.8
AutoDDPM	89.0	85.9	84.2	80.6	27.6	27.0
THOR	89.3	89.1	88.3	84.5	36.2	37.2
SymmDiff (ours)	94.2	91.6	91.1	88.4	43.9	42.7

Values in %. Best results in bold.

Qualitative Localization

Across minor-to-large fractures (Samples 1–3), SymmDiff yields anatomically plausible reconstructions and localizes anomalies consistent with the masks. For the implant case (Sample 4), responses concentrate around the implant with minimal spillover; for the healthy case (Sample 5), activations remain low.

Qualitative pelvic-radiograph localization results — Qualitative localization results on pelvic X-rays. Columns show five different test samples. Rows: input radiograph, corrected reconstruction, binary ground-truth mask, and predicted anomaly heatmap.

Ablations

Each component contributes. Removing reference (symmetry) conditioning causes the largest drop, confirming the side-to-side prior as the strongest stabilizer; ROI-constrained denoising prevents global drift, and multi-scale fusion is necessary for thin fractures that are weak in raw intensity residuals.

Variant	Image-level			Pixel-level
Variant	AUROC	AUPRC	F1_max	AUROC	AUPRC	F1_max
SymmDiff (full)	94.2	91.6	91.1	88.4	43.9	42.7
w/o reference conditioning	89.4	88.2	86.2	79.9	32.4	30.6
w/o random patch masking	93.7	90.8	89.7	86.3	36.0	35.7
w/o ROI-constrained denoising	91.8	90.3	89.8	83.2	35.7	35.1
w/o latent swap	92.6	90.5	89.8	84.5	38.5	37.7
w/o multi-scale fusion (pixel only)	93.9	91.2	90.7	86.6	36.8	35.9

Values in %.

Insights

Insight 1

Contralateral consistency is a strong inductive bias for pelvic UAD — it stabilizes “normality” where global density alone fails.

Insight 2

Symmetry must enter the reverse trajectory, not the post-hoc residual — coupling it during denoising is what prevents anatomical drift.

Insight 3

ROI-gated updates with identity outside the mask make the pipeline robust to imperfect contralateral references.

Insight 4

Multi-scale residual fusion is essential for thin cortical breaks that vanish in any single residual view.

Dataset

Evaluation uses a retrospective pelvic X-ray cohort collected at Chosun University Hospital, South Korea (2015–2025):

Normal radiographs: 1,268, split 80/10/10 into train/val/test — the only data seen during training.
Abnormal cases: 534, held out for evaluation only, with pixel-level fracture masks.
Resolution: images resized to 256 × 256; diffusion runs in a 32 × 32 VAE latent space.
Sampling: T = 1000 training steps; deterministic DDIM with K = 25 steps at inference.

The cohort is retrospective clinical data and is not publicly redistributable. The full training and evaluation code is released so the pipeline can be reproduced on comparable radiograph collections.

Code on GitHub

BibTeX

@inproceedings{rahman2026symmdiff,
  title     = {SymmDiff: Symmetry-Guided Diffusion for Fracture
               Localization in Pelvic Radiographs},
  author    = {Rahman, Abdul and Ghafoor, Afnan and Lee, Bumshik},
  booktitle = {Medical Image Computing and Computer-Assisted
               Intervention -- MICCAI 2026},
  year      = {2026},
  publisher = {Springer},
}