Fully automated deep-learning section-based muscle segmentation from CT images for sarcopenia assessment

To develop a fully automated deep-learning-based approach to measure muscle area for assessing sarcopenia on standard-of-care computed tomography (CT) of the abdomen without any case exclusion criteria, for opportunistic screening for frailty. This ethically approved retrospective study used publicly available and institutional unselected abdominal CT images (n=1,070 training, n=31 testing). The method consisted of two sequential steps: section detection from CT volume followed by muscle segmentation on single-section. Both stages used fully convolutional neural networks (FCNN), based on a UNet-like architecture. Input data consisted of CT volumes with a variety of fields of view, section thicknesses, occlusions, artefacts, and anatomical variations. Output consisted of segmented muscle area on a CT section at the L3 vertebral level. The muscle was segmented into erector spinae, psoas, and rectus abdominus muscle groups. Output was tested against expert manual segmentation. Threefold cross-validation was used to evaluate the model. Section detection cross-validation error was 1.41 ± 5.02 (in sections). Segmentation cross-validation Dice overlaps were 0.97 ± 0.02, 0.95 ± 0.04, and 0.94 ± 0.04 for erector spinae, psoas, and rectus abdominus, respectively, and 0.96 ± 0.02 for the combined muscle area, with R2 = 0.95/0.98 for muscle attenuation/area in 28/31 hold-out test cases. No statistical difference was found between the automated output and a second annotator. Fully automated processing took