Improving Automated Detection of Pediatric Wrist Fractures via Cast Suppression

Overview

This study developed and evaluated paired and unpaired deep learning models to suppress cast artefacts in pediatric wrist X-rays, improving automated fracture detection. By generating synthetic paired datasets and controlling training data composition, the authors demonstrated enhanced fracture detection performance and direct pixel-wise evaluation of cast suppression accuracy.

Background

Wrist trauma is a common cause of emergency department visits, with radiography being the standard diagnostic tool. Immobilisation devices like casts often obscure anatomical details, complicating fracture diagnosis and reducing the accuracy of automated detection models. Previous AI-based cast suppression methods using CycleGANs faced challenges including anatomical distortion risks, lack of paired data for direct evaluation, and confounding effects from training data biases. This study aims to overcome these limitations by creating synthetic paired datasets and comparing paired (Pix2Pix) and unpaired (CycleGAN) cast suppression techniques.

Data Highlights

Key Findings

• CycleGAN was used to generate synthetic cast artefacts on cast-free pediatric wrist X-rays, creating paired datasets for supervised training.
• Pix2Pix models with U-Net architectures of varying capacities were trained on these paired datasets for cast suppression.
• Direct pixel-wise evaluation of cast suppression was enabled by the synthetic paired data, overcoming limitations of prior indirect metrics.
• Fracture detection performance improved when using cast suppression as a preprocessing step, with explicit control of cast presence in training data to avoid bias.
• Paired Pix2Pix models demonstrated superior reconstruction accuracy compared to unpaired CycleGAN models.
• The study design ensured no overlap between cast suppression development data and fracture detection training/testing data, preserving evaluation integrity.

Clinical Implications

Implementing AI-based cast suppression techniques can enhance the visibility of anatomical structures in pediatric wrist X-rays, potentially improving the accuracy of automated fracture detection systems. The use of paired synthetic datasets allows for safer and more reliable image enhancement without risking anatomical distortion, supporting better diagnostic workflows in emergency settings.

Conclusion

This study presents a novel approach combining synthetic paired data generation and supervised learning to effectively suppress cast artefacts in pediatric wrist radiographs, thereby improving automated fracture detection. These findings support the integration of cast suppression models into clinical imaging pipelines to enhance diagnostic accuracy.

References

1. Hržić et al. 2023 -- Cast Suppression in Pediatric Wrist X-rays Using CycleGAN
2. Lee et al. 2023 -- CycleGAN-based Splint Suppression Improves Adult Wrist Fracture Detection
3. Norris et al. 2023 -- Enhancing Radiologist Confidence with CycleGAN-generated Wrist X-rays
4. Isola et al. 2017 -- Pix2Pix: Image-to-Image Translation with Conditional Adversarial Networks
5. GRAZPEDWRI-DX Dataset -- Publicly Available Pediatric Wrist X-rays