Hyperparameter-Optimised U-Net for Bone Fracture Segmentation in Musculoskeletal X-Ray Images

Abstract

Bone fracture segmentation in X-ray images is essential for computer-aided musculoskeletal diagnosis, yet segmentation performance is strongly influenced by model configuration and hyperparameter selection. This study develops a hyperparameter-optimised U-Net model for binary segmentation of bone fracture regions in musculoskeletal X-ray images. The model was trained and evaluated using the FracAtlas dataset, which contains 4,083 annotated X-ray images, supplemented with 327 fracture-specific images obtained from Roboflow. Images and masks were harmonised through grayscale conversion, resizing to 256 x 256 pixels, normalisation, binary mask preparation, and augmentation to improve generalisation. Key hyperparameters, including learning rate, batch size, optimiser, number of epochs, loss function, and augmentation strategy, were empirically tuned using validation performance. The final configuration used a batch size of 8, Adam optimiser, learning rate of 0.001, binary cross-entropy loss, and 200 training epochs. Performance was evaluated using accuracy, validation loss, Intersection over Union (IoU), precision and pixel accuracy. The optimised U-Net achieved a validation accuracy of 99.6%, IoU of 0.995, precision of 0.991 and pixel accuracy of 0.993, demonstrating competitive performance compared with existing musculoskeletal segmentation models reported in the literature, including CMSCNet, 3D U-Net, and modified U-Net variants. The findings show that careful hyperparameter tuning, combined with U-Net encoder-decoder reconstruction and skip connections, can improve fracture-region segmentation in X-ray images. Further validation using larger multi-centre datasets and systematic ablation analysis is recommended before clinical deployment.