Enhancing automated fracture detection in paediatric wrist X-rays with paired and unpaired cast suppression methods.

purposeCasts in follow-up wrist X-rays reduce diagnostic image quality, complicating the assessment of fracture healing. This study developed cast suppression models using real unpaired data and synthetic paired data and investigated their impact on automated fracture detection performance in paediatric wrist X-rays.

methodsA cast suppression model was developed using data generated with unpaired image-to-image translation methods. A published CycleGAN model was repurposed to generate a synthetic paired dataset from an institutional collection of 31,001 X-rays. This dataset was used to train Pix2Pix cast suppression models, using three architectures-U-Net 256, 512, and 1024. Models were evaluated using structural similarity index measure (SSIM), mean-squared error (MSE), and peak signal-to-noise ratio (PSNR). A fracture detection model was trained on a publicly available dataset of 20,327 paediatric wrist X-rays. This model was used to evaluate CycleGAN and Pix2Pix-based cast suppression by testing performance across training sets with varying cast prevalence, defined as 100, 50, 25, and 0% of the naturally occurring cast proportion in the training dataset.

resultsWithin Pix2Pix models, the U-Net 1024 configuration performed best across all metrics (SSIM = 0.957, MSE = 22.2, PSNR = 34.7 dB) and was selected for subsequent fracture detection experiments. Cast suppression improved fracture detection only in models trained without any cast exposure, where Pix2Pix preprocessing increased mAP@0.5 by 4%, from 0.823 (0.008) to 0.859 (0.004), but reduced mAP@0.5 by 1-2% and mAP@0.5:0.95 by 3-4% in models that included cast images during training. Pix2Pix consistently outperformed CycleGAN-based cast suppression across both evaluation metrics (mAP@0.5 and mAP@0.5:0.95), with comparisons assessed using bootstrap resampling.

conclusionThe Pix2Pix model outperformed CycleGAN-based cast suppression for fracture detection preprocessing. However, cast suppression only improved performance in models trained without cast images and reduced performance otherwise. These findings indicate that cast suppression effectiveness depends critically on the downstream model's training data composition.