Work

AI-driven MR thigh scan analysis for body composition phenotypic classification of healthy older persons

Year: 2025

Type: article

Abstract: Sarcopenia, characterized by age-related muscle loss and fat infiltration, reduces muscle function in older adults, increasing healthcare costs. Accurate segmentation of thigh muscle and fat compartments, such as Subcutaneous Adipose Tissue (SAT) and Intermuscular Adipose Tissue (IMAT), is vital for muscle health assessment. MRI enables non-invasive evaluation, but manual segmentation is labour-intensive. The study involved analysis of 190 MR scans from older adults from local population, with an average age of 68 and 69.5 % being female, and a mean BMI of 23.75 kg/m². The MR scans were organized in BIDS format and enhanced using a GAN-based model to improve contrast and signal quality before segmentation with an Attention Residual U-Net (ARU-Net). Statistical evaluation of accuracy utilized Dice and Hausdorff metrics. Volumetric features were employed for AWGS19 classification based on gender, using weak classifiers. ARU-Net-based segmentation Dice scores were - IMAT: 0.8416 ± 0.012, Muscle: 0.9806 ± 0.005, and SAT: 0.9527 ± 0.013, with an average Hausdorff distance below 3 mm. Automated segmentation showed a high correlation (R2 > 0.99, p < 0.001) with ground truth. Gender-based classification using simple neural network yielded F1 scores of 0.78 for NSNO, 0.82 for NSO, 0.63 for SNO, and 0.66 for SO in females, and 0.91 for NSNO, 0.87 for NSO, 0.80 for SNO, and 0.79 for SO in males respectively indicating the robustness of features in characterization and classification. The developed AI-framework effectively extracts and classifies thigh volumetric features, serving as quantitative imaging biomarkers for AWGS-19-compliant stratification of elderly subjects with sarcopenia. • GAN enhances MRI thigh scans, improving SNR and CNR (p < 0.001) for better tissue visualization. • Attention Residual U-Net (ARU-Net) achieves high segmentation accuracy (DSI > 0.85) for IMAT, SAT, and muscle. • Automated pipeline reduces thigh segmentation time from hours to ∼50 s per volume. • AWGS-19-compliant classification into NSNO, NSO, SNO, and SO phenotypes with robust neural network performance. • Semi-supervised ground truth generation enables scalable, reproducible sarcopenia assessment in cohort studies.

Source: European Journal of Radiology Artificial Intelligence

DOI: 10.1016/j.ejrai.2025.100049