Automating multisite muscle segmentation in ultrasound via deep learning through a standardized framework for sarcopenia assessment
Dawei Zhang, Chonglin Wu, Yuxuan Na, Wanrui Li, Ka-Shing Lee, Yu Sun, Yongping Zheng
Full text PDF 
( Download count: 19)
AbstractBackground: Nowadays, sarcopenia is widely affecting numerous people in the world, which directly leads to devastating clinical outcomes like catastrophic falls, functional disability in mobility, and increased mortality. Ultrasound (US) plays an important role in sarcopenia assessment for quantitatively and qualitatively detecting muscle performance by muscle thickness (MT) and cross-sectional area (CSA). However, it has been hindered by operator dependency and a lack of standardization, making manual measurements impractical for large-scale screening. Research Aims Our study addresses this clinical bottleneck by adopting image segmentation models to enable deep-learning-based automated ultrasound system that standardizes the assessment of multiple muscle groups across the body.
Methods: We trained the deep learning models (U-Net and nnU-Net) for segmenting five key muscle groups (biceps, triceps, rectus abdominis (RA), rectus femoris(RF), and peroneal longus and brevis (PLPB) from US images acquired from 94 adult participants (young (n=22), middle-aged (n=22), and elder (n=50)). The models’ segmentation performance was evaluated against expert manual annotations using the Dice Similarity Coefficient (Dice). For the predicted dataset, the results are correlated with body composition results and physical performance data.
Results: The deep learning segmentation achieved excellent performance across all muscle groups. The mean Dice were: RF (0.8260), RA (0.8810), biceps (0.8370), triceps (0.8080), and PLPB (0.9200), indicating a high accuracy compared to the ground truth. Furthermore, these AI-derived metrics demonstrated strong clinical validity through high correlations with body composition data. The appendicular muscles has the highest associations found between triceps CSA and MM (r=0.8212), PLPB MT and ASM (r=0.7462), and RA CSA and BMR (r=0.6398).
Conclusions: By applying a deep learning-powered framework, the ultrasound provides accurate, highly efficient segmentation of multiple muscles for sarcopenia assessment, making US a practical method for sarcopenia assessment and enabling its adoption in community-based screening and the longitudinal management of sarcopenia.Keywords: ultrasound, sarcopenia, deep learning, muscle segmentation, muscle thickness
Dawei Zhang, Chonglin Wu, Yuxuan Na, Wanrui Li, Ka-Shing Lee, Yu Sun, Yongping Zheng (2026). Automating multisite muscle segmentation in ultrasound via deep learning through a standardized framework for sarcopenia assessment. Gerontechnology, 25(s),1-12
https://doi.org/10.4017/gt.2026.25.2.1703.6