A dual-contrast deep learning model for automated assessment of sarcopenia using MRI.
Authors
Affiliations (8)
Affiliations (8)
- Department of Biomedical Engineering, University of Memphis, Memphis, USA.
- Department of Biomedical Engineering, University of Houston, Houston, USA.
- Northwell, New Hyde Park, USA.
- Department of Medicine, Division of Hepatology, Zucker School of Medicine at Hofstra/Northwell, Hempstead, USA.
- Memorial Sloan Kettering Cancer Center, New York, USA.
- Department of Biomedical Engineering, University of Memphis, Memphis, USA. [email protected].
- Department of Biomedical Engineering, University of Houston, Houston, USA. [email protected].
- Department of Biomedical Sciences, University of Houston, Houston, USA. [email protected].
Abstract
To develop a dual-contrast deep-learning model for automated muscle segmentation on spin-echo and Dixon magnetic resonance imaging (MRI) for assessment of sarcopenia and intramuscular fat fraction (FF). One-hundred-seventy-eight MRI and liver frailty measurements were consecutively collected from patients requiring liver transplantation. The retrospective study cohort included 148 spin-echo and 90 water images with 73 FF maps. MRI images were acquired using spin-echo and 2-point Dixon sequences at 1.5T and 3.0T. Two single-contrast (spin-echo, water) and one dual-contrast U-Net models were developed for automated segmentation of paraspinal muscles and estimation of skeletal muscle index (SMI) and FF. Segmentation accuracy was evaluated using Dice similarity coefficient, sensitivity and false positive rate. Segmentation metrics for single-contrast and dual-contrast models were compared using Wilcoxon Signed-Rank test. Linear regression and Bland-Altman analyses assessed agreement between manual and model-estimated SMI and FF. One-way ANOVA and Friedman tests compared SMI and FF estimated using manual, single-contrast, and dual-contrast models. The Independent-Sample Mann-Whitney U test was used to compare differences in SMI and FF between frail and non-frail groups. All statistical tests assumed a significance level of 0.05. The dual-contrast model demonstrated improved segmentation performance for spin-echo images (single-contrast mean Dice: 0.85 ± 0.08; dual-contrast mean Dice: 0.88 ± 0.06) and comparable performance for water images (single-contrast mean Dice: 0.90 ± 0.06; dual-contrast mean Dice: 0.89 ± 0.06), along with good agreement for SMI estimation (slopes: 0.75-0.82, R<sup>2</sup> = 0.60-0.89) and strong agreement for FF estimation (slope = 0.97 for both models, R<sup>2</sup> = 0.97 for both models) and low mean bias (- 2.37% to 5.65% for SMI and < 1% for FF) relative to ground-truth measurements. SMI estimates obtained using both manual and model-based segmentation were lower in frail group compared to non-frail cohort (p ≤ 0.004); however, FF estimates showed no significant differences between frail and non-frail groups (p ≥ 0.995). This study demonstrates the feasibility of using a single generalized dual-contrast model for automated assessment of sarcopenia and intramuscular fatty infiltration, hence facilitating clinical adoption.