Objectives: Proton Density (PD)-weighted MRI sequence is particularly effective for detecting shoulder pathologies but, limited in accurately delineating bone structures due to noise and trauma-induced signal blurring. To mitigate this limitation, this study employed a CycleGAN framework to generate synthetic PD-weighted images from T1-weighted MRI scans to enhance the dataset. Methods: A CycleGAN framework was used to generate synthetic PD-weighted images from T1-weighted MRI scans. A total of 1,330 axial PD-weighted MR images, including both original and CycleGAN-augmented images, were employed to train a YOLOv8 model for detecting the humeral head and scapular regions. Results: The YOLOv8 model achieved a detection accuracy of 98.70% 91.20 % for humeral head and for scapula, respectively, with an intersection over Union (IoU) threshold of 0.25. Conclusion: This study demonstrates the potential of integrating CycleGAN and YOLOv8 for enhancing bone structure localization in PD-weighted MRI, particularly in challenging scenarios with noise and ill-defined borders. Keywords: Shoulder MRI bone segmentation, YOLOv8, CycleGAN Data Augmentation
Corresponding Author: