Deep Learning-Based Automated Segmentation of the Suprascapular Nerve and Adjacent Structures in Dynamic Supraclavicular Ultrasound.
Authors
Affiliations (6)
Affiliations (6)
- Institute of Applied Mechanics, College of Engineering, National Taiwan University, Taipei, Taiwan.
- Department of Physical Medicine and Rehabilitation and Community and Geriatric Research Center, National Taiwan University Hospital, Bei-Hu Branch, Taipei, Taiwan; Department of Physical Medicine and Rehabilitation, National Taiwan University College of Medicine, Taipei, Taiwan.
- Department of Physical Medicine and Rehabilitation and Community and Geriatric Research Center, National Taiwan University Hospital, Bei-Hu Branch, Taipei, Taiwan.
- Research Center for Health Science, University of Santo Tomas, Manila, Philippines; Department of Physical Medicine and Rehabilitation, Our Lady of Lourdes Hospital, Manila, Philippines.
- Department of Physical and Rehabilitation Medicine, Hacettepe University Medical School, Ankara, Turkey.
- Department of Physical Medicine and Rehabilitation and Community and Geriatric Research Center, National Taiwan University Hospital, Bei-Hu Branch, Taipei, Taiwan; Department of Physical Medicine and Rehabilitation, National Taiwan University College of Medicine, Taipei, Taiwan; Center for Regional Anesthesia and Pain Medicine, Wanfang Hospital, Taipei Medical University, Taipei, Taiwan. Electronic address: [email protected].
Abstract
The suprascapular nerve (SSN) provides major motor and sensory innervation to the shoulder. Its accurate identification on ultrasound is challenging because of its small size, low contrast and proximity to structures with similar echotexture. This study aimed to develop and evaluate a deep learning-based approach for multi-class segmentation of the SSN and adjacent structures in dynamic supraclavicular ultrasound. Dynamic ultrasound videos (n = 80) from 42 healthy adults were manually annotated for the SSN, brachial plexus, subclavian artery and omohyoid muscle. A Double U-Net architecture was implemented, incorporating a VGG-19 pre-trained encoder in the first stage and a randomly initialized encoder in the second stage, together with Atrous Spatial Pyramid Pooling and Squeeze-and-Excitation blocks. Ablation experiments examined the effects of encoder choice, architectural design, annotation strategy, loss function and output weighting. Segmentation performance was evaluated using Dice similarity coefficients on an independent test set. The Double U-Net significantly outperformed the baseline U-Net across all annotated structures. For SSN segmentation, the mean Dice coefficient improved from 0.51 ± 0.19 to 0.68 ± 0.11 (p = 0.03), with greater stability across test videos. A pre-trained VGG-19 encoder showed higher Dice scores than alternative encoders, without statistical significance (p = 0.12). Ablation analyses confirmed complementary contributions of both stages (p = 0.52). Multi-structure supervision yielded modest benefit, with optimal performance when all four structures were annotated (Dice 0.68 ± 0.11). Focal loss achieved the highest average Dice score, and optimal performance was obtained by emphasizing the final output with limited intermediate supervision (α:β = 0.1:0.9). The proposed Double U-Net enables reliable multi-class segmentation of the SSN and surrounding structures in dynamic supraclavicular ultrasound, supporting its potential clinical application in suprascapular neuropathy.