Fig 1.
Ankle movement tasks used to generate experimental data set of B-mode ultrasound images.
Dynamic B-mode ultrasound data were collected from the human soleus muscle for five ankle movement tasks with varying levels of force and displacement. Tasks were as follows: (a) (red) seated, constrained ankle plantarflexion with cyclic pushes of the foot on the ground and legs held still, (b) (orange) seated, free ankle plantar/dorsiflexion with foot in the air, c) (green) weight-bearing, calf raises with cyclic heel raises while standing in place, d) (blue) walking at 1.5 m/s on a treadmill, and e) (purple) a mix of directed movements while standing in place with foot on the ground and in the air. Tasks were specifically chosen to elicit combinations of high/low muscle fascicle force and displacement to provide a data set for training machine learning (ML) algorithms that had a high degree of variability. Task color coding is consistent throughout the paper.
Fig 2.
Workflow from B-mode ultrasound image acquisition and processing to machine-learning (ML) model training and real-time muscle length tracking.
A) In each frame, a B-mode image containing human soleus muscle fascicle (yellow line) is captured via ultrasound probe. B-C) Training and ground truth muscle fascicle length change data sets are obtained via UltraTrack software. D-E) Images in each frame are cropped and down-sampled using Python code to implement open-source functions. F) A machine learning (ML) model is trained using outputs from C and E. G) Unique images not seen by the ML model in the training stage are fed into the ML model. H) The ML model yields pseudo real-time muscle fascicle length measurements. I) Performance metrics are calculated (e.g., RMS error and correlation), to compare ML-derived measurements to the ground truth.
Table 1.
Machine learning (ML) model comparisons.
Fig 3.
Muscle fascicle length versus time from optimized ML algorithm with direct training across tasks.
Subplots show raw output of muscle fascicle length (mm) versus time (seconds) estimated using B-mode ultrasound images fed through a support vector machine (SVM) machine-learning (ML) model with optimized hyperparameters for a representative participant X during several cycles of a given ankle movement task (e.g., constrained ankle = red (top); mix = purple (bottom)). In each case, the SVM ML output was generated following a training procedure using image data from the same participant X and same task as that used to measure muscle fascicle length in the test participant X (i.e., direct task to task training). Muscle fascicle lengths derived from inputting the same B-mode images into UltraTrack (UT) software (gray) are included as ‘ground-truth’ to give context regarding correlation and RMSE across tasks. Note that scaling of Y-axes differs from panel to panel.
Table 2.
Ankle movement task comparisons for SVM across training schemes.
Fig 4.
Muscle fascicle length versus time from optimized ML algorithm with cross-participant training in the walking task.
Subplots show raw output of muscle fascicle length (mm) versus time (seconds) estimated using B-mode ultrasound images fed through a support vector machine (SVM) machine-learning (ML) model with optimized hyperparameters for a representative subject X during several cycles of the walking task. In each case, the SVM ML output was generated following a training procedure using image data from walking in each of the other participants (e.g., participant #1 = dark blue, participant #5 = light blue) and then used to measure muscle fascicle length in the test participant X (i.e., cross participant training). Muscle fascicle lengths derived from inputting the same B-mode images into UltraTrack (UT) software (gray) are included as ‘ground-truth’ to give context regarding correlation and RMSE across tasks.
Fig 5.
Muscle fascicle length versus time from optimized ML algorithm with cross-task training in the walking task.
Subplots show raw output of muscle fascicle length (mm) versus time (seconds) estimated using B-mode ultrasound images fed through a support vector machine (SVM) machine-learning (ML) model with optimized hyperparameters for a representative participant X during several cycles of a given ankle movement task (e.g., constrained ankle = red (top); mix = purple (bottom)). In each case, the SVM ML output was generated following a training procedure using image data from each of the other tasks (e.g., constrained ankle = red, mix = purple) and then used to measure muscle fascicle length during walking in the test participant X (i.e., cross-task training). Muscle fascicle lengths derived from inputting the same B-mode images into UltraTrack (UT) software (gray) are included as ‘ground-truth’ to give context regarding correlation and RMSE across tasks.