Ultrasound Segmentation of Cervical Muscle during head motion: a Dataset and a Benchmark using Deconvolutional Neural Networks

Abstract

Objectives: To automate online segmentation of cervical muscles from transverse ultrasound (US) images of the human neck during functional head movement. To extend ground-truth labelling methodology beyond dependence upon MRI imaging of static head positions required for application to participants with involuntary movement disorders. Method: We collected sustained sequences (> 3 minutes) of US images of human posterior cervical neck muscles at 25 fps from 28 healthy adults, performing visually-guided pitch and yaw head motions. We sampled 1,100 frames (approx. 40 per participant) spanning the experimental range of head motion. We manually labelled all 1,100 US images and trained deconvolutional neural networks (DCNN) with a spatial SoftMax regression layer to classify every pixel in the full resolution (525x491) US images, as one of 14 classes (10 muscles, ligamentum nuchae, vertebra, skin, background). We investigated ‘MaxOut’ and Exponential Linear unit (ELU) transfer functions and compared with our previous benchmark (analytical shape modelling). Results: These DCNNs showed higher Jaccard Index (53.2%) and lower Hausdorff Distance (5.7 mm) than the previous benchmark (40.5%, 6.2 mm). SoftMax Confidence corresponded with correct classification. ‘MaxOut’ outperformed ELU marginally. Conclusion: The DCNN architecture accommodates challenging images and imperfect manual labels. The SoftMax layer gives user feedback of likely correct classification. The ‘MaxOut’ transfer function benefits from near-linear operation, compatibility with deconvolution operations and the dropout regulariser. Significance: This methodology for labelling ground-truth and training automated labelling networks is applicable for dynamic segmentation of moving muscles and for participants with involuntary movement disorders who cannot remain still.

Estimation of absolute states of human skeletal muscle via standard B-mode ultrasound imaging and deep convolutional neural networks

“…head torque), and predict those signals directly from the image, in such a way that the network learns a spatial localisation mapping from the labels to the image (like class activation mapping; CAM [76]). Combined with an accurate segmentation [7], [77], an activity map of generated head force, could provide estimated musclespecific contribution to gross head rotational force.…”

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

“…specifically developing methods for analysis of the neck muscles [22], [23]. Recently we contributed a dataset, a methodology for labelling training images suitable for participants with involuntary head movement, and a benchmark deep learning method for segmenting the neck muscles [23], [24]. Here we apply those methods to the investigation of cervical dystonia.…”

“…In this study, we collect a dataset of 61 participants (35 cervical dystonia, 26 age matched controls) in standing posture demonstrating their range of head yaw, pitch and roll rotations. We use our established methods [23] to label a sample of 2,000 US images from this dataset and to train, validate and test a deep deconvolution neural network to segment five paired muscles, vertebra, ligamentum nuchae and skin. Our primary interest is to establish whether from a single axial image, neck muscle shape alone is diagnostic of cervical dystonia.…”

“…US images (192 total, ~3 per participant) from the linked US-MRI dataset were labelled manually by annotating the boundaries around ten muscles, vertebra, ligamentum nuchae and skin. As described previously [22], [23], MRI images showing the same cod liver oil capsule marked plane were annotated and registered to the US images to guide annotation of the US images.…”

“…Two thousand US images (~30 per participant) from the linked US-Vicon data, sampling uniformly the range of pitch, yaw and roll head rotations were labelled manually for the same 13 segments as described previously [23].…”