Deep learning algorithm for predicting subacromial motion trajectory: Dynamic shoulder ultrasound analysis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopedics

Background:

• Dynamic shoulder ultrasonography is valuable for assessing subacromial motion and identifying shoulder pathologies.
• Manual landmark identification in ultrasound is labor-intensive and time-consuming.

Purpose of the Study:

• To evaluate the feasibility of a deep learning algorithm for automated subacromial motion metric extraction from dynamic shoulder ultrasound.
• To compare the accuracy of different deep learning models against manual measurements.

Main Methods:

• A deep learning framework, including convolutional neural networks (CNN) and self-transfer learning-based CNN (STL-CNN) with or without autoencoders (AE), was developed.
• The algorithm tracked the greater tubercle's trajectory relative to the lateral acromion during shoulder abduction/adduction in 17 participants.
• Mean Absolute Error (MAE) between algorithm output and manual landmark labeling (ground truth) was the primary outcome measure.

Main Results:

• STL-CNN models (with or without AE) demonstrated significantly lower MAE than standard CNN for horizontal landmark differences.
• STL-CNN also showed improved accuracy for vertical landmark localization compared to CNN.
• The minimal vertical acromiohumeral distance, a key metric, had significantly lower errors (0.002-0.007 cm) with STL-CNN versus CNN (0.081-0.333 cm).

Conclusions:

• A deep learning algorithm, particularly STL-CNN, is feasible for automatically detecting key anatomical landmarks in dynamic shoulder ultrasound.
• The developed framework accurately captures the minimal vertical acromiohumeral distance, crucial for clinical assessment of subacromial motion.
• This automated approach offers a more efficient and accurate method for analyzing shoulder biomechanics in clinical practice.