Abstract
BACKGROUND: For prediction of many types of clinical outcome, the skeletal muscle mass can be used as an independent biomarker. Manual segmentation of the skeletal muscles is time-consuming, therefore we present a deeplearning-based approach for the identification of muscle mass at the L3 level in clinical routine computed tomographic (CT) data. PATIENTS AND METHODS: We conducted a retrospective study of 130 patient datasets. Individual CT slice analysis at the L3 level was fed into a U-Net architecture. As a result, we obtained segmentations of the musculus rectus abdominis, abdominal wall muscles, musculus psoas major, musculus quadratus lumborum and musculus erector spinae in the CT-slice at the L3 level. RESULTS: The Dice score was 0.95±0.02, 0.86±0.12, 0.93±0.05, 0.92±0.05, 0.86±0.08 for the erector spine, rectus, abdominal wall, psoas and quadratus lumborum muscles, respectively. For the overall skeletal muscle mass, the test data achieved a Dice score of 0.95±0.03. CONCLUSION: Our network achieved Dice scores larger than 0.86 for each of the five different muscle types and 0.95 for the overall skeletal muscle mass. The subdivision of muscle types can serve as a basis for obtaining future biomarkers. Our network is publicly available so that it might be beneficial for others to improve the clinical workflow within examination of routine CT scans.