Muscle ultrasonography texture in young, middle-aged, and older people and its association with functional performance: A machine learning-based study.
Authors
Affiliations (5)
Affiliations (5)
- Department of Rehabilitation Medicine, Peking University First Hospital, Beijing, China.
- Laboratory of Precision Opto-mechatronics Technology, School of Instrumentation and Opto-electronics Engineering, Beihang University, Beijing, China.
- Department of Ultrasonic Diagnosis, Peking University First Hospital, Beijing, China.
- Laboratory of Precision Opto-mechatronics Technology, School of Instrumentation and Opto-electronics Engineering, Beihang University, Beijing, China. Electronic address: [email protected].
- Department of Rehabilitation Medicine, Peking University First Hospital, Beijing, China. Electronic address: [email protected].
Abstract
Skeletal muscle deterioration accelerates with age, leading to muscle atrophy and dysfunction. Musculoskeletal ultrasound as a nonradiative, inexpensive, and portable tool, has potential to evaluate age-related muscle changes. This study aims to detect age-related muscle features based on muscle texture analysis and machine learning approach, and to investigate muscle features that are related to function. Healthy adults were recruited and divided into young, middle-aged, and older groups. 3 muscle architecture and 154 texture features were extracted from transverse (T) and longitudinal (L) ultrasonic images of muscle. Reliability analysis of all texture features was performed. Support vector machine recursive feature elimination (SVM-RFE) was utilized to identify age-related muscle parameters. Muscle features as contributors of muscle function measured by 4 performance tests were detected. 113 participants were recruited. The inter-rater and test-retest reliability showed that 44 muscle ultrasonography texture features were reliable. Among them, 9 features including gray variance-T, 0°short run low gray-level emphasis-T, inertia-L, 0°short run emphasis-T, gray variance-L, gray average-L, 0°run percentage-T, 45°energy-T, and 0°low gray-level run emphasis-L could be used to classify young and older groups with the accuracy of 94.04%, and sensitivity of 93.33% by SVM-RFE. In the multivariate analysis, 3 texture features contributed to functional performance. Gradient mean variance has the highest predictive value. Texture analysis combined with machine learning could provide non-invasive biomarkers to classify muscles of young and old individuals, and assist in function prediction.