Negative binomial modeling of musculoskeletal ultrasound grayscale histograms: a three-device comparison and...

Quantitative musculoskeletal ultrasound enables objective assessment of muscle morphology and clinically viable estimates of tissue composition. This approach to accessible biomedical imaging shows promise for aiding practitioners in the assessment of muscle health. However, inter-device variability in grayscale image interpretation hinders the clinical utility of quantitative ultrasound. We developed conversion models to harmonize grayscale histograms of muscle tissue across three ultrasound devices.

A total of 1,368 longitudinal ultrasound images were acquired from six muscle sites in older adult men (n = 19; age 64.3 ± 8.3 years) using three ultrasound devices on default settings. Grayscale histograms were extracted from each muscle image and modelled using zero-inflated negative binomial regression to characterize muscle tissue echogenicity (mean grayscale value, µ) and heterogeneity (dispersion parameter, α). Heteroskedastic linear regression was used to develop device-to-device conversion models for µ and α.

Conversion models achieved high agreement for µ (Spearman ρ up to 0.906, RMSLE as low as 0.097) but showed greater errors for α at extremes. The device possessing the widest dynamic range, the Philips EPIQ ultrasound machine, exhibited the best performance. The device with the narrowest dynamic range, the SonoSite Titan, exhibited the poorest performance. Conversion performance did not differ by muscle site.

We have demonstrated the feasibility of developing robust conversion models to harmonize grayscale histograms across different ultrasound devices, improving the standardization and clinical utility of quantitative musculoskeletal ultrasound. This post-processing approach provides a viable pathway for harmonizing quantitative ultrasound data across devices without raw radiofrequency access.