ARRS 2022 Abstracts


2052. Comparing CT-Based Manual Muscle Attenuation Measures Against a Fully Automated Deep Learning Tool for Sarcopenic Myosteatosis
Authors * Denotes Presenting Author
  1. Silas Pickhardt *; University of Wisconsin
  2. Kevin Franco Valle; University of Wisconsin
  3. Matthew Lee; University of Wisconsin
  4. John Garrett; University of Wisconsin
  5. Meghan Lubner; University of Wisconsin
  6. Ronald Summers; National Institutes of Health Clinical Center
  7. Perry Pickhardt; University of Wisconsin
A validated and fully automated CT-based deep learning muscle tool has been proven highly effective for myosteatosis assessment and predicting adverse clinical outcomes. This study aims to compare simple manual muscle attenuation measures for potential agreement with the automated measure.

Materials and Methods:
A previously validated deep learning tool that automatically segments skeletal muscle at the L3 cross section was applied to non-contrast CT in 778 asymptomatic adults (mean age, 51.9 years; 473F/305M; renal donor evaluation in 394, CRC screening in 384). Automated L3-level mean muscle attenuation (HU) served as the reference standard for comparison with manual region-of-interest (ROI) HU measures acquired from the paraspinal (PA), psoas (PS), and body wall (BW) muscle groups at the L3 level. Individual and combined muscle group attenuation measures were correlated with the automated measure, including detection of sarcopenic myosteatosis at the 30 HU threshold.

Manual ROI HU measures showed stronger positive linear correlations with automated muscle measures for PA (r=0.80) and BW (r=0.74) muscle groups compared with PS (p=0.60). Correlation further increased by averaging PA+BW (r=0.88) or all three muscle groups (PA+PS+BW, r=0.88). Mean L3-level automated muscle attenuation was 24.8?13.7 HU, with 61.3% below the sarcopenic myosteatosis threshold. Mean manual HU measures were consistently higher: 42.2?15.2, 37.8±13.2, and 41.6?11.0 for PA, PA+BW, and PA+PS+BW muscle groups, respectively, corresponding to mean % differences of 45%, 30%, and 31%. When correcting for these % offsets, the manual sensitivity for detecting sarcopenic myosteatosis (against the automated reference standard) increased from 27%, 25%, and 24% to 96%, 89%, and 79% for PA, PA+BW, and PA+PS+BW, respectively. However, specificity decreased from 100% to 42%, 85%, and 92%, respectively. Similarly, correction via overall average HU offsets (or increasing the manual myosteatosis threshold) increased sensitivity to 79%, 89%, and 91%, and decreased specificity to 83%, 85%, and 77% for PA, PA+BW, and PA+PS+BW, respectively. More onerous linear corrections results in similar performance.

Manual L3-level paraspinal and body wall muscle ROI HU measures correlate better with fully automated cross-sectional analysis than psoas evaluation. By accounting for the systematically higher HU values obtained with manual muscle measurement, “on-the-fly” paraspinal assessment for myosteatosis can identify sarcopenic patients at risk during routine CT interpretation.