AI-based automated body composition analysis from whole-body MRI and (PET-) CT: an intermodal comparison from...

To determine intermodal agreement of AI-derived volumetric body composition metrics between [18 F]FDG-PET/CT-derived low-dose CT and whole-body MRI (wb-MRI).

We retrospectively analyzed same-day whole-body low-dose CT, obtained from [¹⁸F]FDG-PET/CT, and wb-MRI from consecutive multiple myeloma patients from the prospective GMMG-HD7- and GMMG-HD8-/DSMM-XIX-trials. An automated deep learning pipeline based on the TotalSegmentator model, and its MRI variant were used to segment skeletal muscle (SM), subcutaneous- (SAT) and abdominal visceral adipose tissue (VAT) from CT and MRI, respectively. Volumetric measures and single-slice estimates were derived. Agreement between CT- and MRI-based metrics was assessed using Pearson r, intraclass correlation coefficients (ICC), and Bland-Altman analyses.

Forty-two multiple myeloma patients with same-day PET/CT and wb-MRI were included. For AI-based volumetric body composition metrics, CT and MRI demonstrated high agreement (ICC > 0.95) across all SM, SAT and VAT. However, Bland-Altmann plots revealed systematic modality-related differences between MRI and CT: both SM (bias: -2582 ml, p < 0.0001) and VAT (bias: -211 ml, p < 0.0001) were significantly lower when extracted from MRI compared to CT, while for SAT, the difference was not statistically significant (bias: -263 ml, p = 0.1629). For AI-based single slice analysis, agreement remained high for SAT and VAT (ICC > 0.97), while being lower for SM (ICC 0.87). Bland-Altman plots showed that SM was systematically and significantly lower when obtained from MRI than from CT (bias: -18.4cm², p < 0.0001).

Automated body composition measures from (PET-)CT and wb-MRI are highly concordant but show consistent intermodal biases that should be considered when pooling multimodal data or applying modality-agnostic thresholds.

Not applicable.