Body Composition Analysis of Computed Tomography Scans in Clinical Populations: The Role of Deep Learning

Abstract

Background: Body composition is increasingly being recognized as an important prognostic factor for health outcomes across cancer, liver cirrhosis, and critically ill patients. Computed tomography (CT) scans, when taken as part of routine care, provide an excellent opportunity to precisely measure the quantity and quality of skeletal muscle and adipose tissue. However, manual analysis of CT scans is costly and time-intensive, limiting the widespread adoption of CT-based measurements of body composition. Summary: Advances in deep learning have demonstrated excellent success in biomedical image analysis. Several recent publications have demonstrated excellent accuracy in comparison to human raters for the measurement of skeletal muscle, visceral adipose, and subcutaneous adipose tissue from the lumbar vertebrae region, indicating that analysis of body composition may be successfully automated using deep neural networks. Key Messages: The high accuracy and drastically improved speed of CT body composition analysis (<1 s/scan for neural networks vs. 15 min/scan for human analysis) suggest that neural networks may aid researchers and clinicians in better understanding the role of body composition in clinical populations by enabling cost-effective, large-scale research studies. As the role of body composition in clinical settings and the field of automated analysis advance, it will be critical to examine how clinicians interact with these systems and to evaluate whether these technologies are beneficial in improving treatment and health outcomes for patients.