Deep learning-based framework for comprehensive quantification of thigh and calf muscles and adipose tissues from MRI

Quantification of muscles and adipose depots is essential for characterising pathological changes in neuromuscular, musculoskeletal, and metabolic diseases. This study presents a deep learning framework for automated comprehensive analysis of muscle and adipose tissue in the lower extremities.

Axial two-point dixon magnetic resonance imaging data from thigh and calf were retrospectively collected from 25 participants (mean age: 40.5 ± 5.86 years; 64% male) from the Asian Indian Prediabetes Study. A 3D Attention-Res-V-Net pipeline was trained on expert-labelled ground truth data. A cascade of Attention-Res-V-Net models was trained to first quantify the entire muscle region and subcutaneous adipose tissue (SAT) in thigh and calf. Then, thigh and calf-specific networks quantified 13 thigh and 9 calf muscles, respectively. Intermuscular (InterMAT) and intramuscular (IntraMAT) adipose tissues were quantified by intensity thresholding fat-only image volumes within muscle-specific segmentation masks. Resulting fat voxels were multiplied by the voxel resolution to obtain adipose tissue volumes, which were evaluated as relative errors against the ground truth volumes.

Whole muscle segmentation achieved mean DSCs of 92 (thigh) and 87% (calf); SAT reached 95%. Muscle-specific DSCs ranged from 76 to 90% (thigh) and 68 to 90% (calf). InterMAT errors were ~ 21% (thigh) and ~ 19% (calf), while IntraMAT errors ranged from 17.4 to 58.8%. In addition, the high-quality, expert-annotated dataset generated in this study will be publicly released to facilitate future research.

The framework advances muscle-fat composition analysis in the lower limbs by enabling granular quantification of individual muscles, SAT, InterMAT, and muscle-specific IntraMAT.