Detection of Muscle and Bone Regions from MRI Images with Deep Learning

Magnetic Resonance Imaging (MRI) is one of the advanced diagnostic tools used to evaluate shoulder pathologies. Proton Density (PD) weighted MRI sequences are highly effective in detecting edema, but it is not always easy to distinguish the anatomical boundaries of bones due to the reduction in signal-to-noise ratio and blurring of traumatized cortices as a result of trauma. Computer-based diagnostic systems require precise segmentation, identification and localization of anatomical regions in medical images. Deep learning-based object detection algorithms have shown remarkable success in real-time object identification and localization. In this study, the YOLOv8 model is used to detect the regions of the humerus, scapula bones and the supraspinatus muscle from the rotator cuff muscles in 190 coronal PD-intensity MRI images. The YOLOv8 configuration achieved a high success rate in detecting the humeral head, the articulating portion of the scapula (glenoid) and the entire supraspinatus muscle with an intersection over union (IoU) of 0.25. Our findings show significant promise for YOLOv8-based detection of the humerus, scapula and supraspinatus regions, especially in the context of PD-weighted images that are affected by both noise and intensity imbalance, a first in this field.