Enhancing fracture detection in wrist radiographs via paired synthetic data generation

Purpose Cast artefacts in follow-up wrist radiographs reduce diagnostic image quality and complicate the assessment of fracture healing, displacement, and complications. This study evaluated whether targeted image augmentation and cast suppression can improve automated fracture detection performance, particularly when training datasets underrepresent cast images. Methods A previously published CycleGAN model was repurposed to generate synthetic paired datasets for Pix2Pix training. These Pix2Pix models learned image-to-image translation from synthetic cast to cast-less images. Fracture detection experiments assessed the impact of CycleGAN-based synthetic data augmentation and cast suppression across training sets with varying cast image proportions. Performance was assessed using mean average precision at intersection over union thresholds (mAP@0.5, mAP@0.5:0.95) with bootstrap resampling. Results Synthetic augmentation with CycleGAN significantly improved fracture detection compared to models trained without any cast images. However, models trained with limited real cast data outperformed those using synthetic augmentation. The Pix2Pix model consistently outperformed CycleGAN-based cast suppression. Cast suppression only improved fracture detection when applied to models that had never been exposed to cast images during training; when applied to models already trained on cast-containing images, suppression decreased accuracy. Conclusion By generating a synthetic paired dataset, we addressed a critical limitation in cast suppression research and enabled more robust and meaningful architecture comparisons. CycleGAN-based augmentation and cast suppression preprocessing improved fracture detection performance only in models that had not been exposed to casts during training, demonstrating their potential value when real cast data are unavailable. However, models trained with real cast data consistently outperformed those trained with synthetic alternatives, highlighting that real cast data remain essential for developing reliable and clinically robust fracture detection models.