CMSCNet: a context based lightweight musculoskeletal ultrasound image segmentation method.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Musculoskeletal ultrasound (MSUS) is vital for diagnosing disorders but manual analysis is time-consuming.
• Image noise complicates MSUS analysis, hindering efficient assessment of muscle structural parameters.
• Current deep learning methods for MSUS segmentation are computationally intensive and not real-time suitable.

Purpose of the Study:

• To develop a lightweight and efficient deep learning framework for automated MSUS image segmentation.
• To reduce computational load and improve real-time applicability of MSUS image analysis.
• To enhance the accuracy of musculoskeletal disorder assessment through automated feature extraction.

Main Methods:

• A context-based lightweight deep learning framework utilizing a modified U-Net architecture.
• Incorporation of multi-layer perception modules to reduce parameters and enhance efficiency.
• Integration of dense atrous convolution and restructured convolution for improved feature extraction and reduced redundancy.

Main Results:

• Achieved 98.7% accuracy, matching U-Net but with 90% fewer parameters.
• Obtained an average Intersection over Union (IoU) of 0.7227.
• Demonstrated superior capture of muscle aponeurosis details and fiber junctions compared to ground truth.

Conclusions:

• The proposed framework enables automatic, rapid, and accurate extraction of penniform muscle morphological features.
• Provides a foundation for enhanced muscle pathology assessment and interventional therapy precision.
• Contributes to improving the scientific rigor of rehabilitation treatment through precise imaging analysis.