Validation and extension of CycleGAN-based cast suppression methods in wrist radiographs

Abstract

Fractures frequently necessitate stabilisation using casts, which may produce artefacts in subsequent radiographs. These artefacts obscure the bone structure and can impede accurate diagnoses. Current imaging techniques frequently retain the cast due to logistical difficulties in its removal, leading to suboptimal image quality. The application of AI, especially generative adversarial networks (GANs), presents an innovative method for mitigating cast defects without requiring paired images. This research expands upon the CycleGAN model, previously employed for similar tasks, to improve its efficacy by integrating a perceptual loss function and a self-attention layer, with the objective of generating cast-free wrist radiographs that are indistinguishable from original images.