SonoMyoNet: A Convolutional Neural Network for Predicting Isometric Force from Highly Sparse Ultrasound Images.

Area of Science:

• Biomedical Engineering
• Musculoskeletal Imaging
• Machine Learning in Healthcare

Background:

• Ultrasound imaging (sonomyography) offers superior muscle activity measurement compared to surface electromyography.
• Quantifying muscle changes from ultrasound data is computationally intensive, often requiring complex feature tracking.
• Existing methods struggle with deep-seated muscles and require extensive data processing.

Purpose of the Study:

• To develop a computationally efficient method for estimating continuous isometric muscle force from sparse ultrasound data.
• To introduce SonoMyoNet, a convolutional neural network (CNN) for muscle force estimation.
• To evaluate the performance of SonoMyoNet using varying numbers of ultrasound scanlines.

Main Methods:

• Utilized an offline regression convolutional neural network (CNN) named SonoMyoNet.
• Trained SonoMyoNet to learn features from sparse, equispaced B-mode ultrasound scanlines.
• Evaluated network performance by simulating different numbers of single-element ultrasound transducers.

Main Results:

• SonoMyoNet accurately estimated continuous isometric force using as few as four scanlines.
• The network demonstrated robustness against speckle noise and variations in scanline positioning.
• Significantly reduced computational load compared to traditional feature tracking algorithms.

Conclusions:

• SonoMyoNet provides an accurate and computationally efficient approach to estimate muscle force from sparse ultrasound images.
• The method holds promise for wearable systems and non-invasive muscle force monitoring.
• Global feature extraction from sparse ultrasound data is effective for muscle force quantification.