Automatic Extraction of Muscle Fascicle Pennation Angle from Raw Ultrasound Data

Abstract

Compact, wearable, wireless ultrasound (US) sensing systems are promising devices for the observation of human muscle dynamics, offering low power, non-invasive continuous monitoring of contracting muscles. Plane wave imaging is an ideal imaging modality to meet the high frame rate requirements of fast contracting muscles. However, the low power requirements of wearable wireless US constrain the data transfer rates from the probe to the host computer, and limit the maximum number of transducer channels, at the same time discouraging beamforming on the probe. Therefore, it is crucial to extract physiological parameters directly from raw US data for applications demanding fast imaging speeds (like monitoring muscles and tendons in motion). Machine Learning (ML) methods can be employed to effectively extract such features. Although a few recent works demonstrated A-mode US for motion and force prediction, the automatic extraction of structural muscle features from raw data is still in its infancy. This paper demonstrates the feasibility of extracting pennation angles from raw US data on a small dataset of contracting medial gastrocnemius muscles. Automatically extracted labels from US images are used as ground truth to train ML algorithms that predict pennation angles directly from raw US data, without the need for image reconstruction. We employ statistical features, Principle Components Analysis (PCA) and Covolutional Autoencoder (AE) for feature extraction and evaluate Random Forest (RF), Gradient Boosting (XGBoost) and Convoltional Neural Network (CNN) as regressors. Experimental results show that the best method (AE + XGBoost) achieves a mean absolute error of ~ 0.43° that is consistent with the variability of the manually annotated pennation angles reported in the literature, with a memory footprint smaller than 400 kB and less than 5 ms execution time.