Improved Fascicle Length Estimates From Ultrasound Using a U-net-LSTM Framework.

Area of Science:

• Biomechanics
• Medical Imaging
• Machine Learning

Background:

• Brightness-mode (B-mode) ultrasound is crucial for in vivo muscle dynamics measurement in assistive devices.
• Automatic fascicle length estimation from B-mode ultrasound has limitations in pixel-wise accuracy during locomotion.
• Existing methods struggle to achieve precise measurements across dynamic movements.

Purpose of the Study:

• To develop an advanced deep learning framework for accurate fascicle length prediction from ultrasound images.
• To improve the accuracy of muscle dynamics measurement for enhanced assistive device control.
• To overcome the limitations of current automatic methods in capturing pixel-wise accuracy during locomotion.

Main Methods:

• A novel U-net-LSTM architecture was developed, integrating U-net's segmentation with LSTM's temporal analysis.
• Semi-manual ground-truth data was generated from 64,849 medial gastrocnemius ultrasound frames for training.
• The proposed U-net-LSTM model was compared against traditional U-net and CNN-LSTM configurations.

Main Results:

• The U-net-LSTM model achieved superior performance compared to U-net and CNN-LSTM.
• Validation accuracy reached 91.4% with a Mean Square Error (MSE) of 0.1±0.03 mm.
• Mean Absolute Error (MAE) was recorded at 0.2±0.05 mm, indicating high precision.

Conclusions:

• The proposed U-net-LSTM framework significantly enhances fascicle length estimation accuracy.
• This method offers a promising solution for real-time, closed-loop wearable control during locomotion.
• The improved accuracy facilitates more effective muscle dynamics measurement for assistive technologies.