Artificial intelligence for body composition and sarcopenia evaluation on computed tomography: A systematic review and meta-analysis

Abstract

Purpose: Tracing muscle groups manually on CT to calculate body composition parameters and diagnose sarcopenia is costly and time consuming. Artificial Intelligence (AI) provides an opportunity to automate this process. In this systematic review, we aimed to assess the performance of CT-based AI segmentation models used for body composition analysis. Method: We systematically searched PubMed (MEDLINE), Embase, Web of Science and Scopus for studies published from January 1, 2011, to May 27, 2021. Studies using AI models for assessment of body composition and sarcopenia on CT scans were included. Excluded were studies that used muscle strength, physical performance data, DXA and MRI. Meta-analysis was conducted on the reported dice similarity coefficient (DSC) and Jaccard similarity coefficient (JSC) of AI models. Results: 284 studies were identified, of which 24 could be included in the systematic review. Among them, 15 were included in the meta-analysis, all of which used deep learning. Deep learning models for skeletal muscle (SM) segmentation performed with a pooled DSC of 0.941 (95 %CI 0.923–0.959) and a pooled JSC of 0.967 (95 %CI 0.949–0.986). Additionally, a pooled DSC of 0.967 (95 %CI 0.958–0.978), 0.963 (95 %CI 0.957–0.969) and 0.970 (95 %CI 0.944–0.996) was observed for segmentation of subcutaneous adipose tissue (SAT), visceral adipose tissue (VAT), and bone, respectively. SM studies suffered from significant publication bias, and heterogeneity among the included studies was considerable. Conclusions: CT-based deep learning models can facilitate the automated segmentation of body composition and aid in sarcopenia diagnosis. More rigorous guidelines and comparative studies are required to assess the efficacy of AI segmentation models before incorporating these into clinical practice.