Artificial Intelligence-Powered Software for Automated Analysis of Skeletal Muscle Ultrasonography

We aimed to develop and validate a software tool powered by artificial intelligence (AI) by measuring its consistency and predictability of expert manual analysis quantifying lower limb muscle ultrasound images.

We retrieved quadriceps (QC) and tibialis anterior (TA) images acquired by portable ultrasound devices from previously published studies of patients with critical illness, lung cancer, and healthy subjects. A blinded researcher randomly selected 90 QC ultrasound images including 30 from adults with critical illness in the intensive care unit or during the recovery phase, 30 from outpatient adults with lung cancer, and 30 from volunteer healthy adults. Additionally, 90 TA ultrasound images were extracted including 60 selected from adults in the ICU and 30 selected from volunteer healthy adults. Using NIH-Image J software, three experts manually analyzed muscle thickness, cross-sectional area, and echointensity of QC muscles and TA. Automated analyses of the same parameters were performed using a custom-built deep-learning AI model (MyoVision-US). The agreement was determined using intraclass correlation coefficients (ICC; 2-way random effects for consistency) and MyoVision-US predictability of manual analysis using adjusted linear regressions (R²).

While manual analysis took experts collectively approximately 24 hours to analyze all 180 images with each image requiring roughly 8 minutes, MyoVision-US took 247 seconds to analyze all 180 images (163 seconds for QC images and 84 seconds for TA), saving roughly 99.8% of the time used for manual analysis. The consistency was excellent for all QC (ICC=0.85–0.99) and TA (ICC=0.93–0.99) parameters. Values were not attenuated when examining each population group separately, showing excellent ICC values for ICU (ICC=0.91–0.99) and lung cancer (ICC=0.95–0.99) images. Predictability of MyoVision-US was strong for all QC (R²=0.87–0.94) and TA parameters (R²=0.71–0.95). Regardless of muscle and population group, the highest and lowest predictability was for echointensity (R²=0.94–0.96) and cross-sectional area (R²=0.85–0.87), respectively, with good to excellent predictability for muscle thickness (R²=0.76–0.99).

Application of AI automating muscle ultrasound analyses showed improvements in speed and strong consistency and predictability compared with manual analysis of muscle thickness, cross-sectional area, and echointensity, even for challenging images from patients in the intensive care unit and with lung cancer.

This study introduces a method to automate the analysis of lower limb muscle ultrasound images, which may be useful for future research and rehabilitation applications in further populations and muscle groups.