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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gerontology

Background:

• Sarcopenia, a muscle disorder associated with aging, poses a significant healthcare challenge.
• Ultrasound imaging is a practical, radiation-free tool for sarcopenia assessment.
• Current AI methods struggle with accuracy due to operator variability and dataset limitations.

Purpose of the Study:

• To develop a robust AI framework for sarcopenia detection using ultrasound.
• To address limitations in generalization across diverse patient subgroups.
• To enhance diagnostic accuracy and reduce bias in AI-assisted sarcopenia screening.

Main Methods:

• A meta-learning framework integrating BMI- and view-aware episodic adaptation.
• Domain-specific self-supervised pretraining using a ResNet-18 encoder on MSK-US data.
• Reformulating sarcopenia detection as a cross-subgroup generalization problem.

Main Results:

• The proposed method achieved a better balance between sensitivity and specificity.
• Significantly reduced subgroup bias compared to standard CNN and meta-learning baselines.
• Grad-CAM analysis showed attention to clinically relevant muscle boundaries across varying patient demographics and scan orientations.

Conclusions:

• Integrating subgroup-aware meta-learning with self-supervised pretraining enhances robustness and interpretability in ultrasound-based sarcopenia detection.
• The framework mimics clinical interpretation by distinguishing shared anatomy while adapting to subgroup differences.
• This approach offers a practical advancement for reliable AI-assisted sarcopenia screening in diverse clinical settings.