Muscle Cross-Sectional Area Segmentation in Transverse Ultrasound Images Using Vision Transformers.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Musculoskeletal ultrasound

Background:

• Automated muscle cross-sectional area measurement is crucial for assessing muscle architecture and quality.
• Muscle echogenicity, derived from segmented areas, correlates with muscle health.
• Accurate quantification is vital for diagnosing and monitoring neuromuscular disorders.

Purpose of the Study:

• To evaluate deep learning models, including convolutional neural networks and vision transformers, for automated muscle cross-sectional area and echogenicity measurement.
• To validate these models on a large, diverse dataset of ultrasound images.
• To explore the potential of deep learning for classifying muscle pathologies based on texture analysis.

Main Methods:

• Utilized a new database of 2005 transverse ultrasound images from 210 subjects with varying conditions.
• Applied state-of-the-art convolutional neural networks and vision transformers for automated segmentation.
• Performed statistical analyses including Bland-Altman, Pearson correlation, and Intraclass Correlation Coefficient (ICC) for validation.

Main Results:

• Deep learning models achieved near-human-level performance in automated measurements.
• Average discrepancy in cross-sectional area was < 38.15 mm², and in echogenicity was 0.88.
• High reliability was confirmed with a 0.97 Pearson correlation and ICC > 0.97, with no systematic errors.

Conclusions:

• Automated measurement of muscle cross-sectional area and echogenicity using deep learning is feasible and reliable.
• The approach demonstrates potential for clinical applications in assessing muscle quality and diagnosing neuromuscular disorders.
• Preliminary findings suggest deep learning can classify muscle pathologies from texture, warranting further investigation.