Sarcopenia, or low muscle mass, is an underdiagnosed disease associated with many chronic diseases. Automatically measuring muscle mass and quality opportunistically with CT scans is one way to identify patients with or at risk of developing sarcopenia. This model automates the selection and segmentation of the erector spinae muscle group at axial level T12, a muscle group frequently implicated in this disease.
Loss of muscle mass and quality is an indicator of general health and body composition, predictive of negative outcomes and mortality. Automated measurement may provide prognostication and guide therapy.
Non-contrast CT (chest, abdomen/pelvis)
Information on training data
Trained on 328 public low-dose chest CT
scans from the National Lung Screening Trial (NLST) and 258 internal abdomen and pelvis CT scans of healthy kidney donors.
Model performance metrics
Dice scores of 0.92 +/- 0.03 for the NLST dataset and of 0.93 mm +/- 0.03 for internal dataset, showing no significant difference.
(1) Mean absolute error on T12 axial slice selection (in mm) between predicted and reference on NLST and internal datasets.
(2) Dice score between predicted segmentation and manual annotations on NLST and internal datasets.
(3a&b) Bland-Altman analysis and comparison of muscle quantitative imaging biomarkers (mQIBs) yielded by predicted vs manual segmentations on NLST and internal datasets.
See reference publication for more information.