Automated MRI body composition segmentation correlates strongly with CT but shows wide AVF agreement limits

A deep learning algorithm for MRI-based abdominal body composition achieved Dice scores of 0.862-0.923, and strong correlation (r>0.95) with CT, but wide 95% limits of agreement for visceral fat (-8.0% to 35.9%) indicate need for modality-specific norms.

• The algorithm achieved high Dice scores on external MRI dataset: 0.862 (visceral fat), 0.923 (subcutaneous fat), 0.920 (muscle).
• Cross-modality agreement with CT was limited for visceral fat, with 95% limits of agreement of -8.0% to 35.9% in 3D, highlighting systematic differences.
• The model was trained on 105 whole-body dual-echo MRI scans and externally tested on 67 scans, but the CT comparison cohort was small (n=54).

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.