Automated analysis of neck muscle boundaries for cervical dystonia using ultrasound imaging and deep learning

Objective: To provide an automated visualization, pattern analysis and classification of neck muscle boundaries comparing cervical dystonia with healthy controls. Methods: We recorded transverse cervical ultrasound (US) images and whole-body motion analysis of sixty-one standing participants (35 cervical dystonia, 26 age matched controls). We manually annotated 3,100 images sampling a range of pitch, yaw, and roll head movements (2,000 this dataset, 1,100 previous dataset of 28 healthy participants), and trained, validated and tested a deep residual deconvolutional neural network (DCNN) to classify pixels to 13 categories (five paired neck muscles, skin, ligamentum nuchae, vertebra). For all participants in their normal standing posture, we generated held out DCNN segmented US images, extracted segmentation boundaries, transformed the boundaries to principal components and clustered the dystonia participants into two groups using their component scores. Results: For all segments, metrics of segmentation accuracy were Jaccard Index (50±21%) and Hausdorff Distance (5.8±4 mm). For the four deep muscle layers, their boundaries were used to predict central injection sites for each muscle with average precision 94±2%. Using 10-fold cross-validation to select predictive components, linear analysis of component scores identified two significant eigen-patterns discriminating Dystonia from Controls and classifying group membership (Dystonia1, Dystonia2, Control) correctly at 93.4%. Motion analysis showed the groups differed significantly according to head yaw-rotation posture and truncal posture. Conclusion: Neck muscle shape alone discriminates dystonia from healthy controls. Significance: Using deep learning, US imaging allows online, automated visualization, pattern analysis, diagnostic classification of cervical dystonia and segmentation of individual muscles for targeted injection.