Deep Learning Approaches for Automatic Segmentation of Shoulder Bones Using Multi-Sequence MRI

Abstract Shoulder pain is a common complaint in clinical practice, often requiring MRI for diagnosis and treatment planning. Accurate segmentation of shoulder bones from MRI scans is crucial for 3D anatomical assessment. Deep learning (DL) models have shown promise in automating segmentation, yet previous studies have primarily focused on the humerus and scapula, with limited investigation of the clavicle. This study compares DL models for automatic segmentation of the humerus, scapula, and clavicle from MRI. We explore multiple MRI sequences, including PDW SPAIR (coronal and sagittal) and PD-TSE, to assess their impact on segmentation accuracy. DL architectures, including 3D U-Net and 3D Attention U-Net, were trained, and their hyperparameters were optimized. Performance was assessed using the Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and volume variation. Additionally, morphological measurements were analyzed to quantify segmentation errors. The best segmentation performance was achieved using a combination of coronal and sagittal PDW SPAIR sequences with Attention U-Net, yielding DSC values of 0.878 (clavicle), 0.843 (scapula), and 0.931 (humerus), and minimal HD. Our study highlights the importance of multimodal MRI input and model optimization for shoulder segmentation. The integration of morphological assessments enhances clinical applicability, providing reliable automated segmentation for surgical planning and medical training.