Using Deep Learning-Based Methods for Automated Segmentation of Soft Tissues from Shoulder Ultrasound Images

Shoulder pain and injuries present significant challenges to researchers and clinicians for diagnosing underlying structural changes due to the complexity of the shoulder joint. Ultrasonography has been used for diagnosing shoulder impairment and can non-invasively assess structures and mechanical properties of the shoulder. However, the complexity of the shoulder structures often results in diagnostic difficulties and misdiagnosis. Although deep learning of artificial intelligence has been applied in various biomedical imaging, the adoption of deep learning techniques in the segmentation of musculoskeletal ultrasound images, especially the shoulder, is limited. This study addresses this gap by assessing the effectiveness of 3 deep learning models, U-Net, Mask R-CNN, and DeepLab V3+, for the segmentation of soft tissues from shoulder ultrasound images. We collected 721 images from the shoulder area of 17 healthy adults, including the anterior deltoid, medial deltoid, posterior deltoid, and supraspinatus. We employed a combination of three augmentation methods (elastic transform, horizontal flip, and shift scale rotate) to enhance the dataset. The mixed augmentation strategy resulted in U-Net outperforming Mast R-CNN and DeepLab 3+ with a mean Average Precision (mAP) of 78-81%, a mean Intersection over Union (mIoU) of 81-87%, and Recall and Precision values between 91-94% and 87-91%, respectively. The effective use of deep learning methods could assist clinicians on assessing shoulder structures from ultrasound images.