Subgroup-Based Meta-Learning with Domain-Specific Self-Supervised Learning for Sarcopenia Detection from Musculoskeletal Ultrasound.

Sarcopenia is a progressive muscle disorder linked to aging, frailty, and increased healthcare burden. While ultrasound imaging offers a practical and radiation-free tool for assessment, its diagnostic accuracy is limited by operator variability and inconsistent interpretation across patients. Existing deep learning approaches often fail to generalize under small and heterogeneous datasets. We reformulate sarcopenia detection as a cross-subgroup generalization problem and propose a clinically grounded meta-learning framework that integrates BMI- and view-aware episodic adaptation with domain-specific self-supervised pretraining. The model learns to distinguish shared anatomical structures while adapting to subgroup-specific differences, mimicking the way clinicians interpret muscle patterns across patient types. A ResNet-18 encoder pretrained on external MSK-US data provides robust anatomical priors, while subgroup-aware episodic training enhances cross-patient consistency. Experiments on a patient-level ultrasound dataset show that our method achieves higher balance between sensitivity and specificity and significantly reduces subgroup bias compared with standard CNN and meta-learning baselines. Grad-CAM analyses reveal attention to clinically relevant muscle boundaries across varying BMI and scan orientations. This study demonstrates that integrating subgroup-aware meta-learning with domain-specific self-supervised pretraining can enhance robustness and interpretability in ultrasound-based sarcopenia detection, offering a practical step toward reliable AI-assisted screening in heterogeneous clinical settings.