Automated analysis of abdominal body composition using MRI: algorithm development and validation via CT comparison.

AIM: To develop a deep learning-based algorithm for automated segmentation of MRI data for abdominal body composition analysis and to validate its effectiveness against CT-based analysis. MATERIALS AND METHODS: We developed a 3D nnU-Net-based deep learning algorithm to segment three key body components in MRI data: abdominal visceral fat [AVF], abdominal subcutaneous fat [ASF], and skeletal muscle [SM]. The model was trained on 105 whole-body dual-echo MRI scans, and its segmentation performance was evaluated using the dice similarity coefficient (DSC) on an external dataset of 67 abdominal MRI scans. In a cohort of 54 patients with both MRI and CT exams, cross-modality correlation and agreement were assessed between in-phase MRI measurements (from our algorithm) and CT measurements (from a commercial tool) using Pearson's correlation, intraclass correlation coefficient (ICC), and Bland-Altman analysis for both 3D volume and 2D area (at the auto-detected L3 level). RESULTS: The algorithm achieved mean DSCs of 0.862, 0.923, and 0.920 for in-phase and 0.823, 0.913, and 0.906 for opposed-phase 3D segmentation of the AVF, ASF, and SM on the external MRI dataset. MRI-based measurements demonstrated a strong correlation (r>0.95, p<0.001) and excellent overall agreement (ICC>0.95) with CT-based measurements in both 3D and 2D approaches. However, absolute agreement between modalities was less favorable, particularly for AVF, with wide 95% limits of agreement: -8.0%-35.9% for 3D and -43.7%-65.6% for 2D. CONCLUSION: Our algorithm enables accurate MRI-based 3D segmentation of abdominal body composition; however, modality-specific reference values are needed for clinical application given the systematic differences between MRI and CT measurements.