Annotation-Efficient Individual Muscle Segmentation from Dixon MRI: A Multi-Dataset Benchmark of Pretraining and...

Measuring individual muscles and their fatty infiltration on Dixon MRI can support sarcopenia assessment, but manual annotation is time-consuming and annotated datasets are often small. This study investigated accurate individual-muscle segmentation with limited labelled data. Three datasets were used: an abdominal CT dataset of 300 subjects with adipose-tissue labels, a lower-limb Dixon MRI dataset of 25 subjects with 13 thigh and 9 calf muscles, and an independent cohort of 66 subjects for external testing. Network architecture, encoder pretraining, a Dixon-derived fat-fraction input, and a shared CT-MRI encoder were compared. A 2.5D model, subject-level cross-validation, zero-shot transfer, label-free domain adaptation, and few-shot adaptation were also evaluated. The best models reached Dice $0.947\pm 0.017$ for CT adipose tissue and $0.925\pm 0.010$ for thigh muscle; the best 2.5D model reached 0.928, while a reported 3D Attention-Res-V-Net reached 0.857 under a different protocol. Encoder pretraining helped most at the smallest budgets, with ImageNet performing best and comparably to CT-based pretraining at eight subjects. A shared CT-MRI encoder retained accuracy with $44\%$ fewer parameters, and thigh-muscle volumes agreed closely with reference measurements (ICC $(2,1)=0.998$ ). On the external cohort, fine-tuning raised Dice from 0.11 zero-shot to 0.61, 0.76, and 0.88 with four, eight, and 46 labelled target subjects. The fat-fraction input improved segmentation only when very few subjects (two or four) were labelled. Encoder pretraining and efficient model sharing help when annotated MRI data are scarce. A small amount of target-domain annotation substantially improves cross-cohort transfer, and repeated subject-level validation is important for reliable evaluation.