Cast suppression in radiographs by generative adversarial networks

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Pediatric Orthopedics

Background:

• Casts are essential for immobilizing pediatric extremities during healing.
• Cast superimpositions in wrist radiographs can obscure critical diagnostic details.
• Accurate visualization of bone structures is vital for pediatric fracture assessment.

Purpose of the Study:

• To develop and evaluate a CycleGAN-based method for suppressing cast artifacts in pediatric wrist radiographs.
• To assess the performance of different CycleGAN architectures and input sizes for cast removal.
• To enable clearer visualization of underlying bone structures in pediatric wrist X-rays.

Main Methods:

• Retrospective review of 9672 pediatric wrist radiographs, divided into casted and castless groups.
• Application of CycleGAN models, including U-Net and ResNet architectures, with varying input sizes (256, 512, 1024 pixels).
• Quantitative and qualitative evaluation of cast suppression effectiveness.

Main Results:

• A U-Net based CycleGAN architecture with 512-pixel input demonstrated superior performance (P <= .001).
• The CycleGAN models successfully suppressed cast superimpositions in pediatric wrist radiographs.
• The method facilitates the development of software tools for enhanced radiological viewing.

Conclusions:

• CycleGAN is an effective deep learning approach for removing casts from pediatric wrist X-rays.
• This technology can significantly improve the diagnostic quality of pediatric wrist radiography.
• The developed method supports the creation of advanced radiology software for clinical use.