Artificial Intelligence-Assisted Ultrasound for Sarcopenia Assessment: A Scoping Review.
Authors
Affiliations (3)
Affiliations (3)
- Department of Nursing, Shuguang Hospital Affiliated to Shanghai University of Traditional Chinese Medicine, Shanghai, China.
- Shanghai University of Traditional Chinese Medicine, Shanghai, China.
- Department of Ultrasound, Shuguang Hospital Affiliated to Shanghai University of Traditional Chinese Medicine, Shanghai, China.
Abstract
Muscle ultrasound offers a portable, non-invasive, and cost-effective option for sarcopenia assessment, but operator dependency limits its standardization. Artificial intelligence (AI) may help automate and improve ultrasound-based muscle evaluation. This scoping review aims to map the current landscape of AI-assisted ultrasound for sarcopenia assessment, including research trends, methods, clinical uses and performance, and to suggest future work. Following Joanna Briggs Institute guidelines and the PRISMA-ScR checklist, we searched PubMed, Embase, Web of Science, the Cochrane Library, CINAHL, ProQuest, CNKI, WanFang Data, Yiigle, VIP, and Sinomed up to July 25, 2026. Two researchers independently selected studies and extracted data. Thirty studies were included. This field is developing rapidly, with 76.7% of studies published in the last 3 years, primarily from Spain and China. Model development and validation studies (43.3%) were most common. U-Net and its variants were the predominant AI architectures, mainly for image segmentation. B-mode ultrasound with high-frequency linear probes was frequently used. The rectus femoris was the most assessed muscle for sarcopenia (73.3%). AI-integrated models for sarcopenia diagnosis achieved AUCs of 0.640-0.950. Reported Dice coefficients for muscle segmentation ranged from 0.73 to 0.99. In-house datasets were used in 66.7% of the studies, and global classification labels were the most frequent annotation strategy (46.7%). AI-assisted ultrasound shows potential for objective and scalable sarcopenia assessment but remains early-stage. Operator dependency in image acquisition has not been fully resolved. Further investigation should determine the added value of integrating multimodal imaging, combining diverse AI techniques, and evaluating multiple muscle groups. Future efforts should deepen image-processing strategies to extract clinically relevant features and facilitate standardized, scalable implementation in clinical practice.