{
  "abstract": "Introduction Sarcopenia and combined sarcopenia-obesity are clinically relevant, highly prevalent, and associated with poor patient outcomes. Artificial intelligence (AI) tools are increasingly being developed and applied to routine imaging, offering a scalable approach to automated sarcopenia detection. We aimed to summarise the diagnostic accuracy of AI tools in detecting sarcopenia and combined sarcopenia-obesity from computed tomography (CT) or magnetic resonance imaging (MRI) scans.Methods This systematic review and meta-analysis protocol was registered on PROSPERO. Studies were identified that evaluated AI tools for detecting sarcopenia and sarcopenia-obesity combined from CT or MRI in human participants.Titles and abstracts were screened, followed by full-text review. Whenever possible, data were extracted from external or independent test groups. Random-effects meta-analyses were performed for outcomes reported by ≥2 studies, including Dice similarity coefficients for skeletal muscle, visceral adipose tissue (VAT) and subcutaneous adipose tissue (SAT), a spatial overlap metric ranging from 0 (no agreement) to 1 (perfect agreement). Pooled estimates were reported with 95% confidence intervals and between-study variance (tau2). Further performance metrics were synthesised narratively.Results Of 47 studies, 23 were included. One study used MRI; all others were based on CT. Most studies used fully automated deep learning-based segmentation at L3 vertebral level, with sarcopenia defined using established muscle cross-sectional area or index thresholds. Reference standards predominantly consisted of manual or semi-automated expert segmentation.Meta-analysis demonstrated high agreement between AI-derived and reference measurements. Pooled Dice similarity coefficients were 0.960 (95% CI 0.960-0.961; tau2 0.0004) for muscle, 0.965 (95% CI 0.956-0.975; tau2 0.0001) for VAT and 0.975 (95% CI 0.963-0.986; tau2 0.0001) for SAT. Fisher Z-transformed meta-analysis produced comparable pooled estimates (muscle Z-Dice 1.923[95% CI 1.922-1.924; tau2 0.0488], VAT 1.939[95% CI 1.873-2.006; tau2 0.0053], SAT 2.250[95% CI 1.992-2.509; tau2 0.0866]).One study reported threshold-based sarcopenia classification using AI-derived muscle measurements (AUC 0.874; sensitivity 0.718; specificity 0.876).Conclusions AI models show high agreement with reference standards for automated measurements of muscle and adipose tissue from CT and MRI. Although the studies included varied in model design, training data set characteristics, and imaging type, the absolute variation in performance across studies was small, as indicated by the low tau 2 values. Future work should focus on broad external validation and on comparisons to functional, clinically relevant reference standards to facilitate progress towards clinical integration.",
  "authors": [
    {
      "affiliations": [
        "Sandwell And West Birmingham NHS Trust, Worcester, United Kingdom"
      ],
      "name": "Zoheb Anwar Hashim"
    },
    {
      "affiliations": [
        "Sandwell And West Birmingham NHS Trust, Worcester, United Kingdom"
      ],
      "name": "Fumi Varyani"
    },
    {
      "affiliations": [
        "The Royal Wolverhampton NHS Trust, Wolverhampton, United Kingdom"
      ],
      "name": "Sherif Abdelbadiee"
    },
    {
      "affiliations": [
        "Worcestershire Acute Hospitals NHS Trust, Worcester, United Kingdom"
      ],
      "name": "Rasha Shereef"
    },
    {
      "affiliations": [
        "The Royal Wolverhampton NHS Trust, Wolverhampton, United Kingdom"
      ],
      "name": "David Rosewarne"
    },
    {
      "affiliations": [
        "The Royal Wolverhampton NHS Trust, Wolverhampton, United Kingdom"
      ],
      "name": "Philip Harvey"
    }
  ],
  "title": "P371 Diagnostic accuracy of artificial intelligence for CT and MRI detection of sarcopenia and sarcopenic obesity: systematic review and meta-analysis",
  "uid": "ce0a0309-5886-5339-8d7c-499e05cd902d"
}