DeepACSA

DeepACSA segments anatomical cross-sectional area (ACSA) in panoramic ultrasound images of human lower limb muscles to enable quantitative assessment of muscle size.

Key Features:

• Automated segmentation: Fully automated segmentation of ACSA in panoramic ultrasound images targeting rectus femoris (RF), vastus lateralis (VL), gastrocnemius medialis (GM), and gastrocnemius lateralis (GL).
• Deep learning methodology: Convolutional neural networks trained on labeled ultrasound data.
• Training dataset: Model training used 1,772 ultrasound images from 153 participants acquired with three different devices by experienced operators.
• Accuracy and reliability: Reported intra-class correlation coefficients of 0.96 for RF, 0.94 for VL, and 0.97 for GM/GL, with mean differences and standard errors comparable to manual segmentation.
• Image quality dependence: Prediction accuracy is contingent on high-quality ultrasound images, with increased likelihood of inaccurate predictions for low-quality images.

Scientific Applications:

• Muscle function assessment: Quantitative ACSA measurements support evaluation of muscle function and classification of muscular disorder severity and treatment response.
• Research on muscle conditions: Enables analysis of large ultrasound datasets for studies of physical performance, frailty, sarcopenia, and other muscle-related conditions.
• Clinical evaluation when MRI is unavailable: Facilitates muscle size assessment in settings where magnetic resonance imaging is impractical, such as intensive care units.

Methodology:

Convolutional neural networks were trained on 1,772 panoramic ultrasound images from 153 participants acquired with three devices by experienced operators for automated segmentation of ACSA in RF, VL, GM, and GL.